Computers and Electrical Engineering 117 (2024) 109233

Contents lists available at ScienceDirect

Computers and Electrical Engineering

journal homepage: www.elsevier.com/locate/compeleceng

Malware detection for mobile computing using secure and

privacy-preserving machine learning approaches: A comprehensive
survey
Faria Nawshin a,b , Radwa Gad b , Devrim Unal a , Abdulla Khalid Al-Ali b ,
Ponnuthurai N. Suganthan a ,∗
a KINDI Computing Research Center, College of Engineering, Qatar University, Doha, Qatar
b
Department of Computer Science & Engineering, College of Engineering, Qatar University, Doha, Qatar

ARTICLE INFO ABSTRACT

Keywords: Mobile devices have become an essential element in our day-to-day lives. The chances of mobile
Mobile malware analysis attacks are rapidly increasing with the growing use of mobile devices. Exploiting vulnerabilities
Privacy-preserving machine-learning from devices as well as stealing personal information, are the principal targets of the attackers.
Secure machine-learning
Researchers are also developing various techniques for detecting and analyzing mobile malware
Mobile security attacks
to overcome these issues. As new malware gets introduced frequently by malware developers,
Federated learning
Mobile vulnerabilities
it is very challenging to come up with comprehensive algorithms to detect this malware. There
are many machine-learning and deep-learning algorithms have been developed by researchers.
The accuracy of these models largely depends on the size and quality of the training dataset.
Training the model with a diversified dataset is necessary to predict new malware accurately.
However, this training process may raise the issue of privacy loss due to the disclosure of
sensitive information of the users. Researchers have proposed various techniques to mitigate
this issue, such as differential privacy, homomorphic encryption, and federated learning. This
survey paper explores the significance of applying federated learning to the mobile operating
systems, contrasting traditional machine learning and deep learning approaches for mobile
malware detection. We delve into the unique challenges and opportunities of the architecture
of in-built mobile operating systems and their implications for user privacy and security.
Moreover, we assess the risks associated with federated learning in real-life applications and
recommend strategies for developing a secure federated learning framework in the domain of
mobile malware detection.

1. Introduction

This survey paper analyzes the in-built security mechanisms of the mobile operating systems and makes a comparative discussion
between Android and iOS in terms of malware threats. After that, the methods of the conventional machine learning or deep learning
models to identify the malware attacks are explained along with the security challenges of these techniques. To mitigate the issue
of privacy leakage of conventional machine learning models, this survey paper explores several techniques for privacy preservation
of user-sensitive data along with the advantages and drawbacks of these methods. Because of the dynamic and pervasive nature

∗ Corresponding author.
E-mail addresses: fnawshin@qu.edu.qa (F. Nawshin), rg2110609@qu.edu.qa (R. Gad), dunal@qu.edu.qa (D. Unal), abdulla.alali@qu.edu.qa (A.K. Al-Ali),
p.n.suganthan@qu.edu.qa (P.N. Suganthan).

https://doi.org/10.1016/j.compeleceng.2024.109233
Received 12 February 2024; Received in revised form 23 March 2024; Accepted 2 April 2024
Available online 11 April 2024
0045-7906/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Fig. 1. Mobile subscriptions by technology (Billion) [1].

of the malware applications, our survey paper not only analyzes the machine learning or deep learning models but also focuses
on the significance of federated learning in augmenting privacy-preserving mechanisms in Android malware analysis. This survey
paper provides a comprehensive review of how federated learning is incorporated into mobile security strategies. Besides, how
this integration is better at protecting user privacy is highlighted in the paper while effectively detecting malware. However, as the
malware attackers develop different innovative approaches to attack the federated learning models, this survey provides an overview
of the recent security attacks of federated learning and the possible defense mechanisms to detect or mitigate those attacks. Finally,
this paper outlines future research directions considering the measures found in the recent studies for developing a robust and
secured federated learning solution in Android malware detection.
Due to the growing number of mobile phone users, mobile application developers are developing many eye-catching applications.
People use mobile phones for various aspects of daily life, including answering calls, attending online meetings, location navigation
while driving, etc. According to a statistical report by Ericsson [1], Fig. 1 shows mobile subscriptions by technology (billion), and
it is noticeable that there was a sharp increase in mobile subscriptions in 2022. Mobile phones are assumed to replace computers,
as they can perform almost all tasks. More than 4 million applications are available in Google Play and the Apple Store, according
to the report of [2,3]. Malicious application developers are also taking advantage of this surge in usage, and their targets have
shifted from computers to mobile phones to hamper people’s security. According to another statistical report generated by AV-TEST
Institute [4], every day, more than 450,000 new malicious applications (malware) and Potentially Unwanted Applications (PUA)
are registered.
Smartphones have become unavoidable devices worldwide, serving as tools for managing daily activities. However, storing
sensitive and confidential information, such as personal images, bank account details, and different website account credentials,
increases the security threats to mobile phones. In some cases, developers who do not follow the rules and regulations of the
applications’ security framework can also cause attacks. Tangari et al. [5] showed how health applications in mobile phones can
serve as entry points for attackers. He et al. [6] demonstrated that different cryptocurrency wallet applications are more vulnerable
to security attacks.
Bergadano et al. [7] proposed a solution for detecting malicious applications on both Android and iOS (iPhone operating system)
platforms. Additionally, Weichbroth et al. [8] explained the best practices that should be followed to protect the applications. In
their research, they observed the behavior of common mobile security threats. The existing methods and algorithms are always
insufficient to detect new malicious applications.
Garg et al. [9] compared the likelihood of mobile malware attacks between iOS and Android operating systems. Their study
revealed that the Android operating system is more susceptible to security threats than iOS. One key factor contributing to this
difference is that iOS strictly prohibits third-party application developers/providers from uploading their applications directly to
the Apple Store. Instead, iOS developers must submit their applications to the official application market known as the App Store.
Only after undergoing a digital signing process are these applications published in the App Store. Although Android and iOS
have integrated security features to prevent these threats, as discussed in Section 2.1.2, attackers find vulnerabilities and gain
unauthorized access to mobile devices, allowing them to carry out malicious activities. New malicious applications are uploaded to
Google Play daily, emphasizing the urgent need for robust malware detection algorithms.
This study aims to bridge the gap in existing security measures by leveraging machine learning techniques to enhance mobile
malware detection while prioritizing the protection of user privacy. By focusing on the vulnerabilities of the Android platform

2
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Fig. 2. Organization of the paper.

compared to iOS and reviewing current detection methods alongside emerging threats, we emphasize the necessity of our research.
This paper distinguishes the novelties from existing surveys by focusing on the integration of federated learning within the Android
operating system that still remains largely unexplored in current literature. Our analysis compares conventional machine learning
and deep learning models in identifying malware applications. Additionally, we include the solution mechanisms for detecting
specific malware categories, such as adware, backdoors, trojans, riskware, ransomware, and more, considering the design of different
attack models. Unlike previous surveys on federated learning or mobile security, our work delves into the distinct challenges and
opportunities presented by the system and security architecture of Android and user base. We also address the challenges of data
and model heterogeneity in federated learning systems and discuss the mechanisms of overcoming these found in recent papers.
Moreover, we offer a comprehensive analysis of the potentiality of federated learning to enhance privacy and security for Android
devices which is further supported by a detailed examination of current methodologies, their limitations, and future directions.
This approach will provide novel insights and practical recommendations tailored to developers, researchers and other stakeholders
interested in leveraging federated learning to address Android-specific security concerns. The main contributions of this article are
summarized as follows:

1. We provide a detailed analysis of the application and security architectures of the Android and iOS platforms. Outlining how
applications are structured and secured to offer foundational insights into the mechanisms used by these operating systems
to protect against security threats and ensure user data privacy.
2. We analyze conventional machine learning algorithms for mobile malware detection and discuss the privacy and security
threats associated with these methods. We also describe different security attacks on conventional machine learning models
along with the design of the attack models and defense mechanisms found in recent papers. We further introduce advanced
techniques for privacy-preserving and secure machine learning.
3. We present the challenges of data and model heterogeneity in Federated Learning systems and discuss strategies found in the
recent studies to overcome these.
4. We highlight the potential security threats of Federated Learning when implemented in real-world applications and present
the strategies of secure and privacy-preserving approaches to integrate with Federated Learning.
5. We propose the development of secure Federated Learning by synthesizing recent research, outlining the challenges faced,
and suggesting directions for future research to enhance security and privacy in mobile computing.

Paper Organization: Section 2 explains security and privacy problems in mobile computing. Secure and Privacy-Preserving
Machine Learning is discussed in Section 3. Section 4 discusses the existing work on conventional machine learning for mobile
malware detection. Furthermore, a brief explanation of Federated Learning with malware detection is described in Section 5. Finally,
Section 6 discusses the research challenges and future work, and Section 7 concludes the article. The organization of the paper is
shown in Fig. 2.

2. Security and privacy problems in mobile computing

The Android operating system allows the installation of applications from unverified third-party markets, which can pose a
potential security risk. Furthermore, the Android source code is openly released and accessible to all, whereas, in contrast, the iOS
source code remains undisclosed. In the iOS environment, developers are granted access to develop their own applications, while

3
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

unauthorized access is strictly restricted [9]. The attackers did not spare the iOS platform despite using a closed operating system.
Most previous and existing research works focused on Android malware classification and detection. In contrast, researchers are
currently trying to detect and classify malware for iOS, but they are not proven as solid methods. These research gaps nowadays
influence attackers to attack in iOS platform [10].

2.1. Application & security architecture of mobile computing

This section discusses the application architecture in Section 2.1.1 and security architecture in Section 2.1.2 of both Android
and iOS operating systems.

2.1.1. Application structure

The application structure comprises the organization and structure of application components, such as its main components, file
arrangement, and interconnectivity.

Android
The description of application metadata, required permissions, external libraries, and platform requirements are mentioned in
android.manifest file [11]. An Android application is available in (Android Package Kit) APK form. APK file consists of lib, assets,
resource files, manifest, res, and Dalvik bytecode. These files also serve as the sources for static features. The feature vector’s
extraction method depends on each file’s role.

• Permission Features: All the permissions that an application requires to run are mentioned in AndroidManifest.xml file and
these permissions can be used as a static feature to detect whether an application is malicious or benign [12].
• Component Features: The AndroidManifest.xml registers four fundamental components found in classes.dex and AndroidMan-
ifest.xml files. classes.dex file stores system calls that are needed to declare and create those components [13].
• Intent Features: Intents transmit messages between components and intents are found in classes.dex and AndroidManifest.xml
files. With the combination of intents and components, a feature vector can be made which helps in the two components’
association analysis process. [14,15].
• Constant String Features: The strings which are defined by the developer are stored in strings.xml and are defined by smali
code and are stored in .dex files. String frequencies and content can be extracted to analyze the application characteristics [16].
• Resource File Features: Layout files, images, and sounds from res and asset directories are not modified by malicious codes
and this can be served as a static feature [17,18].
• API Features and Opcode: To generate the detection features, the frequency of API calls and Dalvik Opcodes can be considered
to analyze the coding habit of the developers. It can be significantly differentiated between Malware and Benign application
by counting the frequency of API calls and opcodes [17].
• Native Code Features: Native instructions are used by many malware applications from .so files to do malicious activities
which makes it hard to decompile from the compiled code. The invocation frequency of system call and the extraction of arm
opcode frequency help to detect malware applications [17].
• Data Flow and Control Flow Graph Features: The relationship between invocation instruction and the data flow directions can
help to generate a data flow graph and control flow graph. Vector representation can be generated from these graphs which
helps the detection work [17].

The attackers target this application structure and analyze the underlying components for the purpose of exploiting vulnerabilities
in different features to cause harm to the devices or users. The approaches that attackers follow to perform security attacks by using
static features are discussed below.

• Permission Abuse: Malicious applications often request excessive permissions while installing [19]. Alenezi et al. [20] proposed
a solution on how to select required permissions for an application. Sharmeen et al. [21] proposed a framework using deep
learning and semi-supervised techniques to identify malware behavior by analyzing permissions and escalations that achieved
99.024% accuracy to find out new malware.
• Component Hijacking: AndroidManifest.xml gets manipulated and the attackers either replace or hijack the legit components
with a malicious one in order to steal sensitive information of the users. Wu et al. [22] proposed a solution named SCLib to
overcome the component hijacking issue. SCLib is a secure component library designed to provide in-app mandatory access
control (MAC) for Android applications, thereby defending against component hijacking attacks. It defines and enforces a set
of MAC policies within applications to prevent unauthorized component access that addresses weaknesses of the system.
• Intent Spoofing: The exchange of intents between components is manipulated by malicious applications for the purpose of
redirecting data to malicious destinations. By performing this attack, the attackers can gain access to the system and carry out
unauthorized actions [23].
• Code Injection: This involves inserting malicious codes into the executable files of an application such as the classes.dex file.
Choi et al. [24] investigated the underlying conditions responsible for remote code injection attacks.

4
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

• Resource Tampering: The resource files such as images, and layout files are modified by the attackers. They insert malicious
codes into the application which leads to presenting misleading content to the users. Cai et al. [25] suggested solutions for
resource tampering attacks, such as ensuring application checking for camera access and alerting users if it is unavailable
which secures camera usage for foreground applications, and prevents unauthorized touchscreen interactions by detecting
overlay attacks.
• API Abuse: Malware attackers target different APIs also for abuse and to gain unauthorized access to the system and manipulate
the functionalities of the system. Ruggia et al. [26] discussed how mobile banking applications are susceptible to phishing
attacks through the abuse of ‘‘inotify’’ APIs which is a mechanism for monitoring file system events. They demonstrated that
over 10,000 malware samples identified in 2021 exploited ‘‘inotify’’ APIs for malicious purposes. Ito et al. [27] proposed a
framework for the detection of accessing private information by abusing APIs that are not related to the functionalities of the
applications.
• Native Code Exploitation: Native libraries or native codes can be modified by security attackers by inserting malicious native
code to affect the functionalities of the system. Fedler et al. [28] proposed a technique for controlling the native code execution
which can mitigate this issue.
• Dynamic Code Loading: This technique allows an application to load and execute code from external sources at runtime rather
than including it at compile time. Attackers can exploit this feature to bypass static analysis tools and load malicious code
after an application has been installed or execute arbitrary code without the user’s knowledge. Albanese et al. [29] explored
how attackers exploit dynamic loading and dynamic compilation techniques to evade Android’s security mechanisms.

iOS
The configuration file of iOS is called Info.plist and consists of the settings of the applications and metadata. Some of the
components of this file are discussed below.

• Metadata: This involves information consisting of metadata about the applications, such as version number, bundle identifier
of the application, display name, and more. The applications are identified by the iOS based on metadata [30].
• Permissions: The permissions such as location, camera, microphone, and more that are required to install and execute an
application are specified in info.plist file [31].
• Localization: Localized versions of both application metadata and configuration details of the applications for different
languages are mentioned in info.plist. It helps to display the application’s language based on the users’ geographical location
[32].

2.1.2. Security architecture

The security architecture of Android and iOS, the two major mobile operating systems, is designed to ensure the confidentiality,
integrity, and availability of user data and protect against various security threats. This section discusses the security architecture
of both Android and iOS.
Android
Android, a Linux kernel-based operating system developed by Google, incorporates APIs, Java-compatible libraries, and C
programming language middleware. Android utilizes the Dalvik VM, which executes multiple sandboxes simultaneously [33],
ensuring the isolation of Android applications based on the structure discussed in 2.1.1. To safeguard user privacy and secure
application interactions, the Android operating system includes built-in security measures, which are detailed below.

• Shared ID: Shared ID enables data sharing among application components. Applications require the same digital certificate to
obtain a shared user ID and run within the same process.
• Permissions: Permissions are categorized into three groups based on the type of user data access required during run-time:

– Normal Permission: It allows access to user data outside the sandbox with minimal risk. Explicit user approval is not
required, but users can review permissions post-installation.
– Dangerous Permission: It grants access to critical user data like location, camera, microphone, etc. Users must grant these
permissions at application runtime due to potential risks.
– Signature Permission: It is required for applications to sign the same digital certificate and acquire a shared user ID.
Signature permissions are granted by the Android system during installation [34].

iOS
iOS provides default hardware and software-level security features to prevent security attacks:

• System Security: iOS ensures security through encryption, application sandboxing, and isolation. By encrypting the entire
system by default, enforcing strict application sandboxing rules, and utilizing secure boot processes, iOS creates a secure
environment resistant to unauthorized access and attacks [35].
• Data Security: IOS employs encryption for data at rest and in transit to safeguard user data, followed by regular security checks
and updates. Application sandboxing also plays a crucial role in data security by preventing applications from accessing data
outside their specified sandbox without explicit permission [36]

5
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Fig. 3. Common malware types with families [40,41].

• Application Security: iOS adopts a layered approach to application security and provides developers with security APIs and
frameworks to integrate robust security features directly into their applications. This includes secure storage, encryption, and
secure network communications. Besides, the application review process by Apple assures that all applications available on
the App Store adhere to strict security and privacy guidelines that protect users from malware and privacy breaches [37].
• Network & Internet Security: iOS restricts application access to background communication using encryption protocols,
firewalls, and private addresses. This ensures secure data transmission and prevents network-related attacks [38].

2.2. What is mobile malware

Mobile malware is any malicious code that either behaves as a separate application or sometimes injects malicious codes into
the source code of legitimate apps and converts them to mobile malware. Whatever the way is, the ultimate goal of mobile
malware is stealing personal information, reading/writing call logs/contacts, recording call conversations, or gaining full access
control to the mobile phone or locking that smartphone completely. In the last decade, mobile application usage has increased
tremendously, providing various facilities for business and personal use. Malicious applications often disguise themselves as benign
applications [39]. Though Android and iOS apply permission-based access control to regulate which application can access which
resources, applications still declare some additional permissions that do not directly relate to the core functionality. These are called
over-privileged applications. After that, users grant all the permissions required to run an application without explicitly checking,
intensifying the risk of private information leakage. This way, malicious applications are successfully installed on the user’s mobile
and execute harmful activities.

2.3. Mobile malware categories

Understanding the different types of malware is vital to invent new solutions and algorithms. This section discusses some common
categories of mobile malware, including Adware, Backdoor, Ransomware, Riskware, Trojans, Spyware, and more, together with the
detection algorithms found in the existing articles. Fig. 3 shows malware categories with their families according to [40,41]

Adware
Adware performs malicious activities by showing flashing and lucrative advertisements on the user’s mobile screen. When a
user clicks on those malicious advertisements, the malicious application developer generates a certain revenue. Suresh et al. [42]
proposed a machine-learning-based detection algorithm to detect Android adware. They used static and dynamic features for this
classification and obtained 85% accuracy when considering only static features and 76% accuracy when considering only dynamic
features. Then they combined static and dynamic features for classification and got 84% accuracy. Ndagi et al. [43] performed
a comparative analysis of adware classification using different machine-learning algorithms. They found Random Forest was the
best classifier among several, which provides an accuracy of 98% and a false positive rate of 1.7%. Bagui et al. [44] analyzed the
individual adware families using the dataset CICAndMal2017. They used dynamic features for the classification algorithm in their
research.

6
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Backdoor
Backdoor attacks open hidden doors for attackers to enter smartphones. In this attack, the attackers can control the device
remotely without having the mobile phones on hand. Backdoor attacks become successful by bypassing the authentication processes
and getting access to smartphones. Attackers are very intelligent in embedding malicious codes in legitimate applications, and these
malicious codes are executed only when a certain condition is fulfilled. Sometimes backdoor attack is linked with other malware-type
attacks [45]. When a mobile user clicks on any advertisement link, some malicious codes are embedded into the system. The main
purpose of a backdoor attack is to receive/record call logs, send/receive SMS, steal personal information, use space in memory, and
many more. Cho et al. [46] developed a backdoor attack to unlock Android phones without the user’s knowledge. Yao et al. [47]
proposed a solution for identifying real-time passive backdoor behaviors by attackers. Li et al. [48] again proposed another backdoor
attack to beat machine-learning-based algorithms, and they were successful for 99% out of a total of 750 malware samples to bypass
all the authentication mechanisms. Zhang et al. [49] developed a transferable backdoor attack that targets CNN models for malware
detection.

File infector
File infector is a type of malware that is found to be attached to the APK files. The information on how to run or execute an
application is described in the APK files. The malicious applications are automatically executed once those malicious APK files are
installed. The phone setting changes, stealing device ID and IMEI number are the target of the attackers [50]. To the best of our
knowledge, we have not found any detection and mitigation techniques for file infector malware in recent studies.

Ransomware
Ransomware is a type of malware where mobile phones are locked by attackers with a key. After paying the attacker a certain
amount, they send the key to unlock the phone. However, in some cases, users might not unlock or get access to mobile phones even
after paying the ransom amount, consequently losing all data stored in the phone. According to New York Times report [51], over
900,000 mobile phones are infected due to the ScarePackage ransomware family. Another report [52] showed that 150 countries
were the victim of WannaCry ransomware, which caused 35000 mobile devices to be compromised. Gharib et al. [53] proposed
a solution using both static and dynamic analysis named DNA-Droid to mitigate ransomware attacks. Sharma et al. [54] analyzed
permissions used by benign and malicious applications and made a comparison between those to propose a solution against this
attack. Amer et al. [55] developed a model for early prediction of ransomware attacks based on behavioral analysis using static and
dynamic features like API calls, system calls, and permissions. The model overcomes challenges by clustering related features and
transforming them into simplified cluster classes because of massive feature sizes and complex associations. The LSTM model, along
with ensemble machine learning models, is used to classify reformed API and system call sequences and Android permissions, and
it showed notable performance and accuracy in detecting ransomware attacks.

Riskware
Riskware malware behaves like a legitimate application that gets installed as a malware application on the user’s device. It
can change the phone’s settings and then redirect the user to the malicious advertisement page. Stealing personal information and
changing network settings are the goals of the attackers. Mahdavifar et al. [56] proposed a solution to detect riskware using deep
learning. They performed dynamic analysis for the experiment and evaluated their model on the CICMalDroid2020 dataset. They
also compared the performance of their model with other machine-learning algorithms including label-propagation (LP). Their model
achieved an F1 Score of 97.84% and a false positive rate of 2.76% which are higher than LP.

Scareware
Scareware is a type of malware used to deceive users through fake advertisements. This malware pretends to be a benign
application that tries to convince users that installing it will protect their phones. However, users are convinced by the advertisement,
and after installing this, it starts to reveal malicious activities such as stealing the user’s location information. Hamid et al. [57]
proposed a framework that can detect a couple of malware such as scareware, ransomware, and goodware. Bagui et al. [58] proposed
a solution for classifying different malware families, including scareware, using Naïve Bayes, Decision Tree, and OneR algorithm
and observed that Decision Tree achieved the highest accuracy.

Spyware
Spyware is a type of malicious software that gets installed in users’ devices without the knowledge of the user. This malware
tries to gain control over the network settings, data, camera, and microphone of a mobile device. The information they receive
from this malicious access is sent to the attacker’s database. Certainly, the attackers use the information for illicit purposes,
including bank fraudulent transactions and ransomware attacks. Pierazzi et al. [59] used both conventional machine-learning and
deep-learning algorithms to make a comparative study between spyware and other malware. Salih et al. [60] created a spyware
application so that they could understand the malicious activities of spyware. Qabalin et al. [61] proposed a new dataset and then
applied some machine-learning algorithms to detect spyware. The dataset includes network traffic data during spyware installation
and was validated using the random forest classification algorithm that showcases good accuracy in both binary and multi-class
classifications.

7
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Table 1
Recent studies for the detection of each malware category.
Authors Malware category Features used Learning model Performance metric
used/Developed
Suresh et al. [42] 2019 Adware AndroidManifest.xml, Network Random forest Accuracy of 84%
traffic flow
Ndagi et al. [43] 2019 Adware – Random forest Accuracy of 98%
Bagui et al. [44] 2021 Adware Network traffic flow Gradient boosting XGB AUC score of 98.71%
Mohammed et al. [69] 2022 Adware Network traffic flow AdStop Accuracy of 98.02%
Moreira et al. [70] 2023 Ransomware PE headers CNN Accuracy of 93.73%
Farnood et al. [71] 2021 Ransomware Opcodes, API calls, System RansomCare Accuracy of 99.24%
resources
Almohaini et al. [72] 2021 Ransomware Reports from antivirus engines, Hybrid Accuracy of 100%
Behavior and functionalities
Su et al. [73] 2018 Ransomware Permission, Text, System Ensemble Accuracy of 99.98%
operation calls
Alsoghyer et al. [74] 2019 Ransomware API packages’ calls API-RDS Accuracy of 97%
Faris et al. [75] 2020 Ransomware API calls and permissions SSA-KELM AUC score of 98%
Gera et al. [76] 2021 Ransomware API calls, System calls, Network J48 Recall of 99%
flows, Permissions
Manavi et al. [77] 2022 Ransomware PE header of executable files Proposed method Recall of 95.11%
Hamid et al. [57] 2017 Scareware, Lock detected, Text detected, Text K-Means clustering Accuracy of 98.12%
Ransomware, Goodware score, Encryption detected

Trojan
Trojan is a type of malware that also acts like a benign application. Trojans can take the form of other categories of malware,
such as trojan-spy, trojan-sms, and trojan-dropper [62]. Abualola et al. [62] designed a malicious Trojan application that changes
the notification settings of WhatsApp, SMS, and Facebook Messenger without the user’s knowledge. For this purpose, the authors
used the notification listeners service. Nurul et al. [63] proposed a solution to detect trojans. They used several machine-learning
algorithms like Random Forest, J45, and Naive Bayes for this detection. Ullah et al. [64] observed the correlation among different
dynamic features to design another framework called TrojanDetector which detects Trojan malware.

Botnets
Botnet malware gets installed in the mobile phone remotely by a botmaster and users do not have an idea or knowledge about
this malware installation in their mobile phone. Eslahi et al. [65] explained this attack and DroidDream, which can have root
access to the mobile phone and how it installs other malicious applications. Hashim et al. [66] outlined a systematic review of
botnet attacks. They described how reverse engineering and static analysis analyze the reasons behind botnet attacks. Suleiman
et al. [67] proposed a solution for the detection of botnet attacks using deep learning and observed that their solution outperformed
conventional machine-learning algorithms. In their research, they used the ISCX dataset [68] and analyzed 6802 mobile applications,
where 1929 applications were botnet type and the rest were benign applications.
Tables 1 and 2 summarize existing work on specific Malware Type Classification.
Researchers keep developing efficient algorithms to detect and prevent malware attacks, and machine learning-based solutions
have achieved promising results. As new malicious applications are getting uploaded in application markets every day and installed
in users’ mobile phones, it is essential to develop machine-learning algorithms to detect these newly spreading malware applications.
However, conventional machine-learning algorithms are prone to different security threats, as discussed in Section 2.4. Conventional
machine-learning algorithms refer to machine-learning models that learn different patterns from the data and then make a prediction
for new unseen data. These algorithms use a centralized server to store the training data and train the global model, while privacy
or security measures are not included in the algorithms.

2.4. Security threats against machine-learning

Machine-learning is a rapidly progressing field that has already proven its tremendous performance in the detection field.
However, the attackers try various ways of attacking techniques to hamper the model’s characteristics, ultimately affecting the
prediction performance and accuracy of the model. The researchers should address the domain of security attacks on conventional
machine-learning algorithms so that efficient security measures are developed to preserve the privacy and security of the training
data, the efficiency of the model, and the reliability of the predictions. The existing security risks in this field have been divided
into three types and discussed below.

1. Data Poisoning: A data poisoning attack is a type of adversarial attack where the attackers insert malicious data samples into
the machine-learning models intentionally so that the training dataset is exploited. The attacker observes the vulnerabilities of
the model during the testing period. As the malicious samples are mixed with the training data, the accuracy and performance

8
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Table 2
Recent studies for the detection of each malware category.
Authors Malware Features used Learning model Performance metric
category used/Developed
Khurram et al. [78] 2011 Adware Sequences of opcodes K-Nearest neighbor AUC score of 94.9%
Fallah et al. [79] 2019 Adware, Network traffic Naïve Bayes Adware (F1 measure of 74%),
Ransomware, Ransomware (F1 measure of
Scareware 73%), Scareware (F1 measure
of 72%)
Bagui et al. [58] 2022 Scareware Number of bytes initially sent Decision Tree Accuracy of 79.5%
back and forth, Packet size
Pierazz et al. [59] 2020 Spyware Permission, Structure, Network Ensemble Late Accuracy of 98.2%
Fusion (ELF)
Qabalin et al. [61] 2022 Spyware Network traffic Random forest Binary classification (Accuracy
of 79%) and Multi-class
classification (Accuracy of
77%)
Nurul et al. [63] 2019 Trojan System call Random forest Accuracy of 81.2%
Li et al. [80] 2022 Trojan Permissions Voting Algorithm Accuracy of 98.3%
Ullah et al. [64] 2022 Trojan App-APIs, App-Services, SVM Accuracy of 96.64%
Network Flow, System Call
Diyana et al. [81] 2020 Trojan Request permission, API call, KNN Accuracy of 97.83%
Service receiver, Network
traffic
Suleiman et al. [67] 2021 Botnets Permissions, API calls, Intents DNN-2-layer-100 Accuracy of 99.1%
Hojjatinia et al. [82] 2020 Botnets Permissions CNN Accuracy of 97.2%
Mohammad et al. [83] 2020 Botnets Network traffic J48 Accuracy of 85%
Wadi et al. [84] 2021 Botnets Permissions Random forest Accuracy of 97.9%
Yerima et al. [85] 2022 Botnets Permissions and intents Extra trees forest Accuracy of 96%

of the model are highly affected. In this case, the model starts making incorrect predictions for the new unseen data. The
attackers try to control the predictions of the model in a way that the malicious data can be predicted as their desired class.
However, it is difficult to detect this attack because it almost behaves like a legitimate one [86]. Additionally, backdoor attacks
can also be executed by the attackers through data poisoning attacks [87]. Yerlikaya et al. [88] analyzed the performance
of six machine-learning algorithms against data poisoning attacks using four datasets. They applied distance-based label-
flipping and random label-flipping to execute the data poisoning attacks. Distance-based label Flipping selects data points
for label flipping based on their distance from the decision boundary and targets those likely to impact the accuracy of the
model significantly. Random Label Flipping changes the labels of randomly selected data points in the training dataset that
potentially degrade the performance of the model by introducing incorrect labels.
2. Model Stealing: The attackers select different chosen inputs with confidence levels and output labels to steal machine-learning
models. This attack is also called model extraction attack [86]. After stealing the model, attackers insert malicious data into
the training data, and then again, they re-train the model with these malicious data, which has a great negative impact on
the performance of the model’s prediction. Reith et al. [89] showed that using only 854 queries that cost $0.09, the machine-
learning model can be stolen. They selected SVM and SVR models for the experiment that were trained using publicly available
machine-learning datasets. Lee et al. [90] used the deceptive perturbations method for protecting the machine-learning models
from model stealing attacks.
3. Model Inversion: A Model inversion attack can also hamper the privacy of the training data. In this attack, the attackers try
to steal sensitive information from training data after querying the model [91]. They use a reverse-engineering technique on
training data after feeding inputs to the model, and then the attackers analyze the outputs. The attackers gain access to the
characteristics of the model, which were not supposed to be disclosed to any unauthorized people. Model inversion attack
is very dangerous for those applications such as health care and finance, which use training data that contain sensitive
information [86]. Fredrikson et al. [92] developed a model inversion attack that revealed confidence values to create
prediction queries to the machine-learning models. They used decision tree and neural networks for the experiment.

3. Secure and privacy-preserving machine learning

After describing the vulnerabilities of conventional machine-learning algorithms to various security threats in Section 2.4, we
emphasize the critical need to shift focus towards secure and privacy-preserving machine learning approaches. Researchers are now
focusing on developing algorithms for privacy and security measures in machine-learning models. Together with showing good
accuracy, algorithms should also follow some privacy and security steps to protect training data and models from security breaches.
Gradually, secure machine-learning and privacy-preserving machine-learning algorithms have been developed in recent years.
Secure machine-learning (Secure ML) in Section 3.1 and Privacy-Preserving machine-learning (Privacy-Preserving ML) in Section 3.2
are explained.

9
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

3.1. Secure ML

Inventing new techniques for securing machine-learning algorithms is an active research area because new threats are getting
discovered frequently. So, researchers also need to come up with new techniques so that they can overcome these challenges. Some
methods that can ensure the security of the machine-learning models are explained below.

3.1.1. Data encryption

Encryption algorithms can be used to protect machine-learning models from various security threats. Data encryption techniques
are helpful in keeping the data confidential and using it for both training and testing purposes. As many machine-learning algorithms
have to deal with sensitive data, the security of users’ sensitive information will be preserved in this way if a data encryption method
is applied. Inference attacks can also be prevented by using encryption mechanisms [93].

3.1.2. Data de-identification

De-identification is defined as the process of identifying and hiding users’ sensitive information to protect their privacy and
maintain integrity as much as possible [94]. After applying this technique, machine-learning algorithms will be secured from attacks
because the attackers cannot identify and exploit the vulnerabilities in the model since the original data are replaced with some
random values. However, the attackers still try to infer general data distributions. Data Poisoning attacks and Model inversion
attacks can also be prevented by using this method because attackers cannot alter the data with other malicious data, which forces
the model to make incorrect prediction decisions [95].

3.1.3. Data access controls

Data access control is another security measure that helps to protect machine-learning algorithms from security attacks. Machine-
learning algorithms often have to deal with sensitive personal or financial information. Applying access control mechanisms is one
of the core techniques for ensuring the security of sensitive information [96]. This technique helps to prevent unauthorized access
by the attackers. Also, some insiders can harm the models or data through social engineering attacks [97]. People accessing some
parts of the system can perform these social engineering attacks to compromise the model. With access control mechanisms, there
are also other techniques to check the log activities that can detect these human attacks.

3.1.4. Data segmentation

In this method, data is divided into segments that have separate access privileges [98]. This can limit the amount of data that can
be comprised. In this way, the attackers will also have access to reduced data access. In the case of sensitive data such as financial
or personal, if data are segmented into smaller parts, then attackers will not be able to execute a successful attack. If an attacker
somehow accesses to the data then the overall impact on the whole data will be less because the attacker will access the segmented
portion of the data instead of the entire dataset. Baracaldo et al. [99] used data segmentation to mitigate poisoning attacks.

3.2. Privacy-preserving ML

Privacy-preserving machine-learning means applying the mechanism of privacy-preservation of the training data in machine-
learning models. It uses a technique that helps prevent data leakage outside of a pre-defined scope of the algorithm. Data are
distributed across multiple parties who do not want to share their sensitive data with others [100]. Differential privacy, k-anonymity,
homomorphic encryption, federated learning, and more are some techniques that help achieve that goal if integrated with a machine-
learning algorithm. These privacy-preserving machine-learning algorithms also possess the same characteristics of conventional
machine-learning, like building a model on training and predicting a class for new data together with the properties of the
privacy-preservation of the dataset.

3.2.1. Differential privacy

Differential privacy is a type of privacy-preserving machine-learning algorithm. This algorithm uses two techniques: centralized
differential privacy (CDP) and local differential privacy (LDP). In CDP, the central node inserts random noise into the private data,
whereas, in LDP, all the individual nodes add noise to their private data and then send it to the server to share. LDP ensures to
provide greater privacy. However, it can affect the accuracy rate of the model [100]. Current research should focus on this issue
and propose a new solution.

3.2.2. Homomorphic encryption

In the homomorphic encryption method, training and computations occur on the encrypted data. There is no need for
decryption for the training purpose. While integrating with machine-learning algorithms, homomorphic encryption helps to perform
computations on users’ sensitive data without exposing it to unauthorized people. It allows both training and evaluating models
on encrypted data. This is quite a helpful technique for privacy preservation as many cases, machine-learning models need to
be built on users’ sensitive and confidential data. There are two types, one is fully homomorphic encryption (FHE), and another is
partially homomorphic encryption (PHE). In FHE, all the computations are performed on encrypted data, while in PHE, only specific
calculation types such as addition or multiplication are allowed [101].

10
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

3.2.3. K-anonymity
In k-Anonymity, identity anonymization of the users’ is applied to the dataset to hide their identity from exploitation. This
technique is also one of the privacy-preserving techniques that help to preserve the privacy and confidentiality of users’ sensitive
training data after integrating it into the machine-learning model. Each data sample must be distinguishable from at least k-1 other
data samples. Generalizing and suppressing are two ways to achieve k-anonymity, making it difficult to find a person from their
data. In the case of age and income-related data, this method groups the dataset into different age ranges and income brackets
to prevent personal identification while allowing other useful information extraction from the dataset. This method is helpful for
dealing with sensitive data like medical, personal, and financial data [102,103].

3.2.4. Secure multi-party computation (SMPC)

In this approach, computations are divided into smaller sub-computations and then each party can perform sub-computations
on each part jointly with their own private data [104]. Without disclosing their private input data samples to others, each party
performs the computations on the functions. The primary focus is to keep the privacy of the input data from being accessible by
others, and then they perform sub-computations separately on the function. The results of all the sub-computations are combined
together, and then the final result is disclosed to all other parties. In this method, no party can access the other party’s individual
input data, which helps preserve the privacy of the users’ confidential information [105].

4. Conventional machine-learning for mobile malware detection

The in-built application and security architecture of both Android and IoS, concept of mobile malware along with different
malware categories are discussed in Sections 2.1–2.3 respectively. As mobile operating systems’ built-in features are insufficient to
protect them from malicious applications, many solutions have been proposed using conventional machine-learning/deep-learning
algorithms in the last few years. Static analysis, dynamic analysis, and hybrid analysis have been found in recent papers for the
detection of mobile malware. These analyses are performed on both benign and malicious applications to observe the differences
between them. This section discusses some of the significant approaches to mobile malware detection using conventional machine-
learning/deep-learning algorithms found in recent studies. Section 4.1 presents mobile malware detection approaches based on
static analysis, Section 4.2 presents based on dynamic analysis and Section 4.3 describes the methods based on hybrid analysis.

4.1. Static analysis-based malware detection methods

In static analysis, the code of an application is analyzed without executing for the detection of mobile malware [106,107].
This type of analysis is faster and includes observing assembly code, binary code, and the source code. It tries to discover the
patterns and signatures of malicious applications, such as function calls or strings of code. However, the static analysis does not
help detect encrypted and dynamically loaded malicious code [108]. The static analysis executed only in special conditions identifies
the malicious codes of applications. Though static analysis is efficient and faster than dynamic analysis, it has some limitations. Static
analysis may not detect new malware correctly. There is a chance of producing false positives if the code of a benign application is
similar to that of known malware [109]. The detailed description of the static features is explained in Section 2.1.1.
Researchers use different subsets of the expansive malware dataset and new malware emerges daily. For the effectiveness
of the malware detection solution, it is necessary to train the model using the most recent malware dataset. For these reasons,
conducting experiments using a unified dataset is not possible. Table 3 provides a summary of various studies on mobile malware
detection using static analysis. The authors employed different features and learning models to analyze datasets consisting of a
varying number of samples. For instance, Suarez-Tangil et al. [110] introduced an Android malware classification system focusing
on fast, accurate detection and resilience against obfuscation. It utilized static analysis to extract obfuscation-invariant features and
identified obfuscation methods used in malware. Their method achieved up to 99.82% accuracy with zero false positives for malware
detection and 99.26% for family identification of obfuscated malware. Mat et al. [111] introduced a Bayesian probability model for
Android malware detection based on permission features extracted via static analysis. It utilized a dataset of 10,000 samples from
AndroZoo and Drebin databases, applied information gain and chi-square algorithms for feature selection, and achieved an accuracy
rate of 91.1%. Their approach signified the effectiveness of permission-based features in identifying malware. Bayazit et al. [112]
presented a deep learning-based Android malware detection system employing static analysis with RNN-based LSTM, BiLSTM, and
GRU algorithms. They used the CICInvesAndMal2019 dataset, which included 8115 static features, and their approach demonstrated
the BiLSTM model that outperformed other models with an accuracy rate of 98.85%. Their work highlighted the effectiveness of
deep learning in detecting malware and classifying complex patterns within static data for accurate malware identification. Yang
et al. [113] introduced a novel Android malware detection method called AMDASE, which leveraged API semantics extraction to
improve malware detection. It focused on API clustering for semantic extraction and optimized call graphs to extract robust API
contextual information. Feature vectors were generated by AMDASE for machine learning classification by summarizing APIs into
cluster centers. Each approach presented in the recent papers followed distinct methodologies with their respective constraints.
Despite showing robustness against malware evolution, their methods face challenges in processing efficiency and scalability due
to their reliance on API semantics extraction in static analysis. Some approaches only focus on permission-based features, which
are limited by the evolving nature of malware threats that make the potential for adversaries to manipulate permissions to evade
detection. These limitations emphasize the ongoing challenges in Android malware detection and emphasize the need for continuous
refinement of detection methodologies to address the dynamic and sophisticated nature of malware threats.

11
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Table 3
Mobile malware detection using static analysis.
Authors Features Learning model Datasets No. of Accuracy
used/Developed samples
Sandeep et al. [114] 2019 APK files/Permissions Random forest Google Play Store, – 94.65%
VirusShare
Singh et al. [115] 2019 Permissions and API calls Linear-Support Vector Drebin 16,100 99.6%
Machine (L-SVM)
Mat et al. [111] 2021 Permissions Naïve Bayes AndroZoo [116] and 10,000 91.1%
Drebin [117]
Syrris et al. [109] 2021 Permissions, receivers, SVM Drebin 1,29,013 99%
services and API calls
Akbar et al. [118] 2022 Permissions SVM, Rotation forest, Google Play Store, Virus 10,000 86.25–89.96%
Naïve Bayes, Random Share [119]
forest
Bayazit et al. [112] 2022 Permission RNN-based LSTM, BiLSTM CICInvesAndMal2019 [14] 8115 98.85%
and GRU
Ibrahim et al. [120] 2022 Permissions, receivers, XGBoost CICMalDroid2020 [121] 17,341 97.3%
services and API calls
Odat et al. [122] 2023 Permissions Random forest Malgenome 3798 98%
Chaudhary et al. [123] 2023 Permissions and intents CNN CICInvesAndMal2019 1584 96.4%
Ding et al. [124] 2023 Bytecode image CNN Drebin 4962 95.1%
Nguyen1 et al. [125] 2023 Permissions, intents, 1D-CNN Drebin & CICMaldroid2020 15,130 & 99.6%
services, and API calls 17,016
Ariffin et al. [126] 2023 Permissions KNN Kaggle Dataset 399 95%
Bakır et al. [127] 2023 API calls, Application ANN-based Auto-encoder, Drebin, Malgenome and 6000 98.33%
metadata, Manifest CNN-based Auto-encoder, Google Play
components VGG19-based Auto-encoder
Yang et al. [113] 2024 API AMDASE Custom Dataset 84,604 96.7%

4.1.1. Obfuscation procedure of static features

The static features are obfuscated by the developers of the malicious applications to a certain level so it becomes difficult to
reverse engineer. The common approaches for static feature obfuscation are discussed below.

1. Rename Identifier: Developers follow some naming rules and conventions while coding and usually follow similar ways for
all the identifiers. Random strings are injected by the malware developers and renamed the identifier in a way that makes
it difficult to understand the logic of the code and reverse engineer. Because of not adding many unrelated nodes to the
control flow graph, it is difficult to detect the obfuscated identifier. Garcia et al. [128] proposed a method to detect malicious
applications in spite of having many obfuscation techniques.
2. Insert Junk Code: Operation instructions (nop), additional register, and jumps can contain junk instructions and statements
by malicious application developers, which increases the code’s complexity and deletes the actual static features. Because of
the limited influence of the junk code on data flow, normal data flow work is observed on source–sink [129].
3. Load Dynamic Code: Remote networks, source files contain additional Dalvik bytecode and native code, which can be loaded
by the Android application. The DexClassLoader function is used to load dynamic code by the malware developers. Sometimes
reflection calls can be originated by using lang/Ljava/reflect/Method, which makes it hard to find the malicious code.
The sensitive functions of the applications can be used to determine whether it is benign or malicious. The dynamic calls
are checked and determined whether it is malicious or benign from the (super control flow graph) sCFG, which Poeplau
et al. [130] designed.

4.2. Dynamic analysis-based malware detection methods

Dynamic analysis is a malware detection technique that monitors the run-time behaviors of applications in either virtual or
real environments. It has been promising because it has the capability to resist code transformation. In dynamic analysis, mobile
malware detection is performed by analyzing the behavior of a program or application while it is running. Monitoring system calls,
network traffic, and file system changes are a few approaches to dynamic analysis. These methods are more powerful than static
analysis because they can identify malicious activity that may not be detected through static analysis, which observes the code of
a program or files without executing it [131]. Emulation, Sandboxing, and Honeypots are some of the dynamic analysis techniques
for malware detection. Safe and isolated environments are chosen to observe the dynamic analysis of malicious applications so that
other benign applications on the device are not affected.
Table 4 presents a summary of studies focusing on mobile malware detection using dynamic analysis. The authors employed
various features and learning models to analyze datasets consisting of different numbers of samples. For instance, Haq et al. [132]

12
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Table 4
Mobile malware detection using dynamic analysis.
Authors Features Learning model Datasets No. of Accuracy
used/Developed samples
Mahindru et al. [136] 2017 APK files/Permissions Naive Bayes (NB), Decision Android Botnet 11,000 67.64–99.7%
Tree, Random Forest (RF), [137], DroidKin
Simple Logistic (SL), k-star [138]
Feng et al. [139] 2018 Permissions, API calls Naive Bayes, KNN AndroZoo & 24,019 94.5%
Drebin
Haq et al. [132] 2021 Permissions, API calls Convolutional neural AndroZoo, 38,842 98.83–99.20%
networks, Bidirectional Long Android
ShortTerm Memory (BiLSTM) Malware Dataset
(AMD)
Sihag et al. [133] 2021 System calls, binder Deep neural network AndroZoo & 13,533 98.08%
calls Drebin
Li et al. [140] 2022 API calls, system calls, LSTM VirusShare 43,007 97.31%
and the relationships
between these calls
Bhat et al. [134] 2023 System calls, binders, Stacked ensemble (Logistic CICMal- 11,598 98.08%
and complex Android regression, Support vector Droid2020
objects machine, Decision Tree and,
Naive Bayes as base and
Random Forest as meta
classifier)
Liu et al. [135] 2024 Graph nodes SeGDroid CICMal2020 and 16,595 98.37%
MalRadar [141] and 2092

proposed a dynamic and robust deep learning-based model for Android malware detection that used a hybrid architecture combining
Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks. Their method emphasized
efficiency in identifying persistent malware across various datasets and achieved superior performance in detection accuracy and
time efficiency compared to recent studies. Sihag et al. [133] proposed a deep learning-based framework for Android malware
detection using dynamic features. Their approach was distinguished by the dependence on behavioral characteristics obtained
through dynamic analysis in an emulated environment which proved highly effective with a detection rate of 98.08% and an
F-measure of 98.84%. Bhat et al. [134] introduced a system call-based Android malware detection approach utilizing ensemble
machine learning techniques. Their approach used system calls, binders, and Android-specific composite behaviors to identify
malicious applications. They employed feature selection to enhance classification efficiency and achieved notable success with a
stacking ensemble method which resulted in a classification accuracy rate of 98.08%. Liu et al. [135] presented a novel Android
malware detection method based on sensitive function call graph (FCG) learning. A graph pruning method was employed in their
approach to focus on security-related API calls that used word2vec and centrality measures for node representation and graph
embeddings through a Graph Convolutional Neural Network (GCN). SeGDroid achieved high detection performance with an F-
score of 98% on malware detection and 96% on malware family classification, which emphasized the ability to capture semantic
knowledge of malicious behavior effectively. However, these techniques have several limitations. Some approaches face challenges
in scaling to process large volumes of applications due to computational complexity and time constraints and difficulties in detecting
entirely new or evolving malware without periodic retraining or updates to include novel malware samples and techniques. Despite
some methods being designed to resist obfuscation, sophisticated adversaries continuously develop new evasion techniques that can
bypass current detection mechanisms.

4.3. Hybrid analysis-based malware detection methods

Hybrid analysis for mobile malware combines both static and dynamic analysis techniques. This technique has been proven to
have greater accuracy in malware detection in recent studies. After selecting an application, it analyzes the code of that application.
Then the application is executed in an isolated environment like sandboxing to protect other applications. As hybrid analysis
uses both techniques together, this technique is considered more effective in malware detection than using those two techniques
separately [142]. However, these methods may also be more resource-intensive and time-consuming because they use static and
dynamic features together.
Table 5 provides a summary of studies on mobile malware detection using hybrid analysis. For instance, Huanran et al. [143]
proposed a hybrid Android malware detection framework that integrated static and dynamic analysis techniques. They utilized
permission sequences and object reference relationships to enhance detection accuracy. By adopting machine learning algorithms,
their approach achieved an F1-measure of 97.5%, which demonstrated effectiveness in resisting permission abuse behaviors and
obfuscation techniques. Mahdavifar et al. [144] introduced a semi-supervised deep learning framework for Android malware
detection utilizing pseudo-label stacked auto-encoders (PLSAE). This hybrid approach combined static and dynamic analysis to
prepare feature vectors and evaluated them in the CICMalDroid2020 dataset. It showed notable improvements over traditional

13
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Table 5
Mobile malware detection using hybrid analysis.
Authors Features Learning model Datasets No. of Accuracy
used/Developed samples
Surendran et al. [147] 2020 API calls, permissions and Ridge regularized Drebin, Androzoo 3300 97%
system calls logistic regression Google Play Store
Hadiprakoso et al. [148] 2020 Manifest permission, API call XGB Malware Genome, 35,000 99%
signatures, intent-filter, Drebin, CICMalDroid
command signature, and
binaries
Lu et al. [149] 2020 API call, intents, permission, DBN-GRU Public malware sharing 13,298 96.82%
dynamic code loading, system websites, Google Play
call store
Dhalaria et al. [150] 2020 API call, intents, permission, Random Forest Available in Github and 3600 98.53%
Command Strings, Kaggle
Cryptographic Operations,
Dynamic Permissions,
Information Leaks, System
Calls
Upadhayay et al. [151] 2021 Permissions, Network Traffic SVM Genome, Drebin, 4900 95.96%
Koodous [152], Google
Play store
Huanran et al. [143] 2022 Permissions, memory TextCNN Google play, VirusShare 21,708 99.8%
Mahdavifar et al. [144] 2022 Intents, Permissions, System PLSAE CICMalDroid2020 17,341 98.28%
calls, Composite behaviors,
Basic binders
Chaganti et al. [142] 2023 API call sequences and PE CNN - 40,000 97%
Headers
Taher et al. [145] 2023 Command strings, API calls, DroidDetectMW Drebin and 10,450 98.1%
intents, and permissions, CICAndMal2017 [153]
dynamic permissions,
cryptographic operations,
system calls, and information
leakage
Dhalaria et al. [146] 2024 - MalDetect Drebin and AndroMD 15,036 98.57% and
and 3547 97.90%

supervised and semi-supervised methods and achieved high accuracy and efficiency in utilizing a significant volume of unlabeled
data alongside a smaller set of labeled samples. Taher et al. [145] proposed a hybrid model that integrated static and dynamic
analysis for Android malware detection. Their proposed model used feature selection strategies to optimize the classification
process and employed a neural network optimized with an enhanced Harris Hawks Optimization (HHO) algorithm. Their approach
significantly improved the accuracy of malware detection and classification and highlighted the importance of hybrid analysis in
combating sophisticated malware. Dhalaria et al. [146] focused on improving Android malware detection through a classifier fusion
approach, using various machine learning techniques and ranking algorithms based on predictive error rates. Drebin and AndroMD
datasets were used in their experiment to demonstrate the effectiveness, which achieved better performance than individual base
classifiers and traditional ensemble learning techniques. Despite achieving high-performance scores, these approaches include several
limitations, such as handling new or unknown malware without frequent retraining, the need for large and diverse datasets for
training, and computational demands for deep learning models and dynamic analysis in real-time scenarios.
The articles we included in this section used conventional machine-learning algorithms that do have in-built integrated privacy
preserving techniques or security measures in the design of the algorithms. Though conventional machine-learning algorithms do not
include security and privacy-preservation techniques by default, several privacy-preserving algorithms such as federated learning,
differential privacy, k-anonymity, or more can be integrated with these algorithms. The techniques for ensuring secure machine-
learning and privacy-preserving machine-learning are discussed in Section 3 and Federated Learning in 5. The model needs to be
trained using a large dataset containing users’ sensitive information to detect mobile malware. So, it is very crucial to integrate
privacy-preserving techniques in the malware detection algorithms, which not only show greater accuracy in malware detection but
also preserve the confidentiality of the data.

5. Federated learning

Federated Learning [154] is a privacy-preserving machine learning technique where the devices train the local models with their
own training samples; no other device can access those data. Section 3 explains other approaches for developing secure ML and
privacy-preserving ML. In Federated Learning, the local models transfer the parameters only to the central server, which is needed to

14
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Fig. 4. General framework of FL [158].

create the global model. The global model is built by aggregating all the parameters sent by all the local devices. After aggregating,
the global model is shared among all the devices to start the process again [155,156]. These processes continue until the desired
accuracy is achieved by the algorithm [157]. Fig. 4 shows the general framework of Federated Learning (FL) [158]. In this method,
the user’s data privacy is preserved by not allowing the data to leave the devices. So, the centralized server for storing all the
training data, like in the conventional machine-learning approach, is not followed here. This section describes the applications of
FL in real-life scenarios along with the solution mechanisms of data and model heterogeneity in FL systems. Section 5.1 outlines
the integration of FL approaches used for mobile malware detection along with the security threats and challenges. After that,
Section 5.2 introduces different solution mechanisms from the literature for developing a secure FL framework.
There are considerable advantages to using FL over conventional machine learning. The privacy of the users’ data is preserved
because the data does not leave the local device. The local devices only share the parameters. The central server also does not
know about the data, and it minimizes the chances of many security risks. As training is performed based on large and diverse data
samples, the model’s accuracy and generalization capability are increased.
Some applications already have been using the FL technique, such as natural language processing [159], image recognition [160],
healthcare [161], the internet of things [162], intrusion detection systems [163], data mining [164] and many more. In the case of
medical data, FL helps to perform training on large and diverse medical datasets from multiple hospitals where any hospital cannot
access another hospital’s data. Same to image recognition, FL can be used to train models on a large set of images within the local
devices, and it preserves the users’ privacy.
Qayyum et al. [165] used clustered FL for the automatic detection of COVID-19. They used two popular datasets to evaluate their
solution. Their solution obtained favorable results for both datasets. According to their result, they achieved 16% and 11% higher
F1-Score for using the FL-based technique than the conventional technique for those two datasets. Alamer et al. [166] developed a
privacy-preserving FL approach called PPFL-SC using IoT devices for advanced malware detection. It incorporated group-oblivious
signcryption to safeguard data privacy and employed an incentive mechanism to enhance IoT device collaboration. This methodology
ensures the security and privacy of data while effectively detecting malware through collective learning from diverse IoT sources.
FL systems are prone to security threats where adversaries try to change the model parameters to hamper the accuracy of the
predictions’ decisions. FL systems are susceptible to data poisoning attacks, eavesdropping attacks, model inversion attacks, and
others [167–169]. Researchers have been developing different algorithms to protect the FL systems from different cyber-attacks and
increase the model’s performance. FL is still a promising technique for several real-world applications, but securing the privacy of
users’ data is of high priority. As FL performs training collaboratively on large datasets [170], developing efficient techniques to
detect those cyber attacks is of utmost importance. Section 5.2 discusses the solution mechanisms to prevent these security attacks
of FL systems.
Addressing Data and Model Heterogeneity in FL Systems
Despite having lots of advantages of using an FL approach, it still has some challenges, such as Data and model heterogene-
ity [171]. As there are many clients and each client can have different types and formats of data, which creates data heterogeneity
issues. To address data and model heterogeneity in the FL systems, Cao et al. [172] proposed a solution using Generative Adversarial
Networks (GANs) within a personalized FL framework called PerFED-GAN. This method utilizes the power of GANs to generate
synthetic data samples that are representative of the client’s unique data distribution. The use of GANs addresses the issue of data
heterogeneity and ensures the privacy of the client’s data because synthetic samples can be shared without exposing the original
data. Furthermore, the solution introduces a co-training mechanism that is designed to ensure that the synthetic data generated

15
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

by the GANs remains aligned with the data distribution of the client. This mechanism considers both the diversity of the client’s
data and model architectures and ensures that the aggregated global model can generalize across the heterogeneity present within
the FL system. After that, Wang et al. [173] incorporated data generators within the FL framework. In this method, training data
is supplemented with a generative model, and its learning process and the storage of the supplementary data are performed on
the server side. Optimizing the global model is constrained by the distance between the training global model and the aggregated
global model’s distribution to ensure effective learning without divergence between real and synthetic data. Li et al. [174] proposed a
solution called DynamicNet that addressed data privacy in heterogeneous environments in the FL system by dynamically adjusting
model aggregation weights and privacy budgets. In this method, privacy budgets are used to determine how much data can be
shared without compromising individual privacy, and aggregation weights are used to influence the importance of data from each
node during model updates. The solution is applied to mobile computing, where nodes may vary widely in terms of data quality
and quantity along with computational power. Chen et al. [175] introduced a decoupled model architecture called FedCMD that
differentiates between cloud and edge computing resources within an FL system. It uses contrastive learning techniques to customize
model architectures for edge nodes that allow for personalized model updates. It enables the model to learn from differences between
nodes and improve the ability to handle heterogeneous data. This solution not only increases model generalization ability across the
network but also overcomes the challenges of data and model heterogeneity. Azimi-Abarghouyi et al. [176] proposed a hierarchical
structure for organizing clients in an FL system. The method is combined with quantization techniques to manage communication
and computation, and clients are grouped based on their similar data characteristics or computational resources. After that, clusters
are formed, and localized model updates are made. This solution addresses statistical heterogeneity by confirming that model updates
are more representative of the data distribution of the network.

5.1. FL for mobile malware detection

Privacy-preserving machine-learning approaches have already been used to detect mobile malware to protect users’ sensitive
information. FL techniques have been used with different machine-learning or deep-learning algorithms to detect mobile malware.
It can not only detect malicious applications with higher accuracy but also preserve the privacy of the training data.
The model can be trained on large and diverse datasets captured from various sources using FL. As mobile malware detection
requires access to sensitive data like system logs, user behavior data, and more, FL techniques play a vital role in this respect. This
technique enables a decentralized approach where the raw data do not leave out of the local devices. The global model is developed
by aggregating the parameter updates from local devices.
The following are the general reasons why to use FL in mobile malware detection based on papers [177–182].
Privacy: People store confidential information on mobile devices such as photos, personal information, bank account details,
locations, addresses, different websites login credentials, and many more. While detecting mobile malware, the applications must
access the users’ devices. As FL systems do not allow leaving data out of the devices, the privacy of these sensitive data is preserved.
Data Diversity: Malicious developers continuously develop malicious applications and make the malware dataset large and
diverse. Training machine-learning models with this diverse dataset is necessary to reduce false positives or negatives. FL helps in
this process of training models with large and diverse datasets.
Real-time Detection: As the FL technique means the model will be trained locally based on the local data, it is more efficient
to detect malware in a particular user’s device based on its own data, ensuring real-time detection of mobile malware.
Resource Efficiency: It ensures resource efficiency because FL systems do not transfer data from devices to a central server. As
only the parameters are transferred, it reduces the computational burden of the central server.
Table 6 shows the research work on Mobile Malware Detection using FL. For instance, Hsu et al. [183] used privacy-preserving
FL to detect Android malware. API calls, permissions, and learned local models have been used in this paper to detect Android
malware. In their experiment, FL has been used with a secure multi-party computation technique and SVM. The work used the
dataset from the Opera Mobile Store. Privacy of users’ data has been ensured in this framework. To the best of our knowledge, this
was the first work on Android malware detection using an FL system. Gálvez et al. [177] proposed a framework called Less is More
(LiM) for the detection and classification of Android malware using FL. The information about the other installed applications in
the mobile phone is kept private by using FL. In this experiment, the authors used a semi-supervised ensemble algorithm to achieve
higher accuracy in both detection and classification. The performance metric showed that the global model achieved an F1 Score
of 95% and local clients achieved 1% false positive in more than 100 applications. They used a dataset consisting of 25,000 benign
applications and 25,000 malicious applications. In the security analysis section, they mentioned that their model is robust and
secure against poisoning and inference attacks. Jiang et al. [178] proposed a new Android malware classification framework named
FedHGCDroid using FL, which detects and classifies Android malware in a privacy-preserved way. The authors used a convolutional
neural network and a graph neural network for the experiment. Though integrating these two neural network algorithms could
classify the Android malware with high accuracy, they introduced the concept of FL to allow collaborative training by several
Android users. This can assure both privacy of the data and the capability of real-time detection. In this work, the Androzoo dataset
was used and their work proved that this method was more adaptable and accurate in the detection and classification of Android
malware than other existing solutions. Chaudhuri et al. [180] proposed a solution named FedAvg using FL for Android malware
detection. In this approach, the global model is developed by merging all the updates from all the local models. This approach has
the limitation that if there is any poor-performing local model, it can hamper the performance of the global model because this
approach gives value equally to all the local models. To overcome this issue, the authors updated their solution using a dynamic
weighted federated averaging (DW-FedAvg) technique in which the local models’ weights are updated based on their performance.

16
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Table 6
Mobile malware detection using FL.
Authors Analysis method Features Learning model Datasets No. of F1 score
samples
Hsu et al. [183] 2020 Static Permissions and API SVM Opera Mobile 61,730 95.92%
Calls Store
Gálvez et al. [177] 2020 Static Permissions, LiM AndroZoo [116] 50,000 95%
Activities, Services,
Intents, Hardware
components,
Broadcast receivers
Jiang et al. [178] 2022 Dynamic Behavioral features FedHGCDroid AndroZoo 70,000 91.25%
Chaudhuri et al. [180] 2023 Dynamic – DW-FedAvg Melgenome, Between Between
Drebin, 3799 and 97.63% and
Kronodroid 78,137 99.18%
[184] and
Tuandromd
[185]
Mahindru et al. [179] 2023 Static Permissions DNNdroid – 100,000 97.8%
Amjath et al. [186] 2024 Static API calls and DotGAT AndroZoo and 2500 90%
Opcodes Maldroid [121]

Fig. 5. Navigating security in FL for mobile malware detection.

In this experiment, Drebin, Tuandromd, Kronodroid, and Melgenome datasets were used, and the performance evaluation showed
that this dynamic version outperformed the traditional one.

Security Threats in FL:

FL systems have found applications across real-world scenarios where training data is confidential and cannot be shared
with external parties. Recent studies have identified several security threats to these systems. Fig. 5 highlights the critical need
for integrating privacy-preserving and secure machine learning techniques with the FL framework in mobile malware detection.
However, FL systems remain vulnerable to security threats even with these integrations. Section 5.2 discusses the strategies for
addressing these vulnerabilities.

1. Data Leakage: The attackers try to infer private data after analyzing gradients of FL. This is called a data leakage attack and
this attack can occur during the gradient-sharing process of FL. Xin et al. [187] developed an advanced data leakage attack
called CAFE. They also proposed solutions to mitigate data leakage attacks.

17
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

2. Byzantine Attacks: According to recent studies, the FL system is prone to Byzantine attacks [188] where the malicious clients
perform these attacks. This is one of the dangerous attacks because it can drop the accuracy of the model from 100% to
0% [189] by updating these parameters just opposite of the linear combination of other normal updates and the accuracy of
the global model drops to 0%. The attackers try to perform this Byzantine attack either on training data or on parameters. If
the attack is on training data then it is also called a data poisoning attack where the attackers manipulate the local training
data which ultimately affects the global model’s accuracy. Backdoor triggers, adding noise, and label flipping are some of
the ways to attack training data. The attackers try to change the local parameters either on the model or gradient so that the
global model receives the corrupted copy of parameters which also affects the accuracy of a global model.
3. Data Poisoning Attacks: In this attack, the model is modified or tampered with by the attackers. The attackers either insert
malicious samples in the training set or tamper with the existing dataset by label-flipping. The accuracy of the model is
affected by this attack [190,191]. Doku et al. [192] proposed a solution for mitigating data poisoning attacks for FL systems.
Psychogyios et al. [193] introduced two types of data poisoning attacks in FL and also proposed a solution to mitigate this
attack based on server-side clean-label training.
4. Model Poisoning Attacks: In this attack, the attackers manipulate the parameters of the chosen model and the global model
starts behaving erroneously. This attack can create the chances of a backdoor attack [190]. According to [191], model
poisoning attacks can be more effective than data poisoning attacks. The attackers often try the bit-flipping method to convert
the model into the trojan-infected model [194]. As the local updates sharing to the central server and then the global model
sharing with local clients are repetitive processes, attackers try to change the parameters of the model in between these
processes.
5. Privacy Breaches: The privacy-preservation of training data is the primary target of the FL system. However, the attackers try
to reveal sensitive information about the users from the local models or during the aggregation of updated parameters [195].
Chai et al. [196] proposed a framework using Hierarchical Blockchain-Enabled FL to mitigate this threat.
6. Communication Attacks: The procedures inside a FL system are iterative where the performance of the global model is
improved after completion of each round. There are a huge number of communication messages that need to be exchanged
between central servers and local clients. The convergence of FL is achieved after running a large number of iterations. During
the exchange of communication messages, the communication channel is prone to many security attacks [188].

5.2. Secure FL

Uprety et al. [197] analyzed the effect of data poisoning attacks on the training samples such as label-flipping attacks. They
proposed a method to mitigate this attack by observing each node’s reputation scores using the beta probability distribution to
identify and filter out malicious nodes and evaluated the model using the MNIST dataset. The approach enhances the robustness of
FL models by improving the trustworthiness of contributions to the global model. Furthermore, it incorporates a blockchain-based
network with FL to prevent denial of service attacks and ensure the privacy of the data and integrity of the system, which is an
advantage of the method. However, this method has several challenges and drawbacks. It introduces computational overhead during
the calculation and updating of reputation scores, which can affect the efficiency of the model. The method considered only one
type of data poisoning attack since there are other types of adversarial attacks, which leads to the adaptability issue of the proposed
solution to various attack vectors.
Unal et al. [198] proposed a solution to mitigate the FL attacks by integrating blockchain-based systems with FL in IoT-based
networks. This integration enhances the privacy-preserving capabilities of FL by ensuring the integrity of the trained models
through blockchain to mitigate model poisoning attacks. The approach used fuzzy hashing to detect anomalies in FL-trained models
against poisoning attacks which added an additional layer of security. However, the solution introduces complexities for integrating
blockchain technology, which could potentially lead to increased computational overhead and latency in processing IoT data. There
is also a scalability challenge by the model because of the reliance on fuzzy hashing and maintenance of an immutable blockchain
record.
Ali et al. [199] proposed a blockchain-based solution to improve the security of FL by showing IoT-based use cases in visualized
and scattered FL environments. The advantages include enhanced privacy and security through decentralization that reduces the
vulnerability to attacks and unauthorized access by distributing data across a network. Additionally, transparency and immutability
of the blockchain network strengthen trust among participants. However, the solution has some drawbacks, such as scalability
issues, extensive computational requirements, potential latency of the blockchain network and the complexity of implementing a
blockchain network with federated learning and significant overheads in terms of energy consumption.
Houda et al. [200] proposed a solution called MiTFed that enables the development of a global model for intrusion detection
jointly by several software-defined networks. They did not share any information regarding model training with one another.
This work also proposed a framework using the SMPC technique for the secure aggregation of updates from the local models.
In addition, a blockchain-based system was also adopted by the authors so that the collaboration can be done in a trustworthy
and decentralized way. They evaluated their solution using the popular network security dataset NSL-KDD. MiTFed addressed the
concerns of distributed denial-of-service (DDoS) attacks and the limitations of traditional intrusion detection systems (IDSs) that
are prone to zero-day attacks due to their reliance on signature-based detection methods. However, the method introduces certain
challenges and drawbacks because of the integration of complex structures such as FL, blockchain, and SMPC, which can lead to
increased computational overhead and latency. Additionally, the effectiveness of the framework depends on the participation and
collaboration of multiple SDN domains, which raises the concern of data privacy.

18
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Liu et al. [201] proposed a solution named BFG for mitigating attacks in decentralized federating learning using blockchain,
differential privacy, and Generative Adversarial Network. The advantages include because of utilizing blockchain, BFG removes the
single point of failure issue inherent in central servers, which enhances the robustness and reliability of the system, and also, the
chances of inference attacks were reduced by 26% in this solution. The integration of differential privacy introduces noise to the
data, which hides the details of individual data samples and strengthens the privacy of the data. However, the method has some
drawbacks, including significant computational resource requirements and performance overheads that can affect the speed of the
learning process of the model.

6. Research challenges and future work

The field of mobile malware detection has several research gaps and future research directions; some of these key areas include:

• Scarcity of Extensive Datasets: The possibility of accurate malware detection depends on the use of real-world extensive
malware datasets. With the increasing number of new malicious applications, training detection algorithms with real-world
and complete datasets become very crucial. After gaining access to these real datasets by the researchers, they can build robust
machine-learning algorithms capable of accurately detecting and classifying new malware.
• Advanced Bypassing Techniques: Malicious application developers are developing various algorithms to enable malware
applications to bypass detection mechanisms and gain access to mobile devices. To address this concern, researchers should
focus on this issue and propose new solutions that can detect these bypassing attempts by malware applications.
• Privacy and Security Concerns: As the accuracy of detecting and classifying mobile malware depends on training machine-
learning models with large and diversified real-world datasets, it arises the issue of disclosing users’ private information.
Though there are already some techniques proposed in recent studies for ensuring privacy-preserving machine-learning, still
it needs to access large data or parameters. Researchers should work on this issue and develop algorithms for mobile malware
detection that can achieve higher accuracy while minimizing the need for a large collection of datasets.

In the past few years, the importance of mobile malware detection has increased significantly. This is due to the intent to
replace laptops/computers with mobile phones for their convenience and the ability to use them for almost all purposes. Different
conventional machine-learning algorithms were developed initially for the detection and classification of malware applications.
These conventional machine-learning algorithms achieved higher accuracy and performance metrics in differentiating between
benign and malicious applications, however, these algorithms fail to preserve the privacy of the users’ data. Hence there are multiple
techniques discovered by the researchers such as differential privacy, k-anonymity, FL, and more which ensure the confidentiality of
users’ data. To the best of our knowledge, only the FL technique has been integrated with machine-learning algorithms by researchers
for the detection and classification of mobile malware while also preserving the privacy of the training dataset.
However, the FL system is prone to various security attacks such as data poisoning attacks, model poisoning attacks, inference
attacks, and more. Researchers have proposed some security measures to mitigate FL attacks. Still, there are some drawbacks
to each approach. Many papers, such as [197–201] have proposed solutions using blockchain. However, blockchain has several
disadvantages, including high implementation costs, environmental impacts, storage problems, and many more. Moreover, while
blockchain is integrated with FL to secure the FL environment, the blockchain network itself is prone to many security threats. The
security measures to protect the blockchain network while integrating with FL have not been addressed in these papers. Future
work should focus on resolving this issue. Furthermore, most of the papers have not evaluated the algorithms using real-world use
cases for ensuring secure FL. The researchers should address this part to develop solutions that can effectively deal with real-world
scenarios. Additionally, future work should include the development of new innovative other privacy-preserving algorithms to detect
mobile malware with higher performance metrics.

7. Conclusion

With the ever-increasing usage of mobile devices, malware developers are developing various malicious applications to cause
harm to devices or users. To overcome this threat, the primary goal is to detect mobile malware. There are many machine-learning
and deep-learning algorithms have been developed by researchers. This paper gives a summary of recent conventional machine-
learning algorithms for this detection. Since the performance of machine-learning models depends on the quality and diversified
training data, this training procedure on a large dataset can lead to the exposure of private information of the users. This paper
reviews some techniques in recent studies on preserving the privacy and security of users’ confidential information while applying
machine-learning algorithms. Furthermore, the paper discusses some challenges associated with preserving privacy and security
in the context of mobile malware detection. It gives constructive directions for future work to develop more robust and secure
algorithms to identify and detect mobile malware threats effectively.

CRediT authorship contribution statement

Faria Nawshin: Conceptualization, Literature search, Writing – original draft, Editing. Radwa Gad: Literature search, Writing –
original draft, Editing. Devrim Unal: Conceptualization, Supervision, Review & editing. Abdulla Khalid Al-Ali: Supervision, Review
& editing. Ponnuthurai N. Suganthan: Supervision, Review & editing.

19
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Data availability

No data was used for the research described in the article.

Acknowledgments

Open Access funding provided by the Qatar National Library. This study has been partially supported by an internal GA Grant
from Qatar University.

References

[1] Ericcson. Ericsson mobility report. 2022, https://www.ericsson.com/4ae28d/assets/local/reports-papers/mobility-report/documents/2022/ericsson-

mobility-report-november-2022.pdf. Accessed on: 2024-01-12.
[2] AppBrain. Number of android applications. 2022, https://www.appbrain.com/stats. Accessed on: 2023-12-10.
[3] Statista. Number of apps available in leading app stores as of 3rd quarter 2022. 2022, https://www.statista.com/statistics/276623/number-of-apps-
available-in-leading-app-stores/. Accessed on: 2023-12-10.
[4] AVTest. Malware. 2022, https://www.av-test.org/en/statistics/malware/. Accessed on: 2023-12-13.
[5] Tangari G, Ikram M, Sentana IWB, Ijaz K, Kaafar MA, Berkovsky S. Analyzing security issues of android mobile health and medical applications. J Am
Med Inform Assoc 2021;28(10):2074–84.
[6] He D, Li S, Li C, Zhu S, Chan S, Min W, Guizani N. Security analysis of cryptocurrency wallets in android-based applications. IEEE Netw 2020;34(6):114–9.
[7] Bergadano F, Boetti M, Cogno F, Costamagna V, Leone M, Evangelisti M. A modular framework for mobile security analysis. Inf Secur J: Glob Pers
2020;29(5):220–43.
[8] Weichbroth P, Łysik Ł. Mobile security: Threats and best practices. Mob Inf Syst 2020;2020:1–15.
[9] Garg S, Baliyan N. Comparative analysis of android and iOS from security viewpoint. Comp Sci Rev 2021;40:100372.
[10] Ciaramella G, Iadarola G, Martinelli F, Mercaldo F, Santone A. A model checking-based approach to malicious family detection in iOS environment.
Procedia Comput Sci 2022;207:1981–91.
[11] Lin J, Liu B, Sadeh N, Hong JI. Modeling {users’} mobile app privacy preferences: restoring usability in a sea of permission settings. 2014, p. 199–212.
[12] Zarni Aung WZ. Permission-based android malware detection. Int J Sci Technol Res 2013;2(3):228–34.
[13] Kim T, Kang B, Rho M, Sezer S, Im EG. A multimodal deep learning method for android malware detection using various features. IEEE Trans Inf
Forensics Secur 2018;14(3):773–88.
[14] Taheri L, Kadir AFA, Lashkari AH. Extensible android malware detection and family classification using network-flows and API-calls. In: 2019 international
carnahan conference on security technology. ICCST, IEEE; 2019, p. 1–8.
[15] Feizollah A, Anuar NB, Salleh R, Suarez-Tangil G, Furnell S. Androdialysis: Analysis of android intent effectiveness in malware detection. Comput Secur
2017;65:121–34.
[16] Wang W, Gao Z, Zhao M, Li Y, Liu J, Zhang X. DroidEnsemble: Detecting android malicious applications with ensemble of string and structural static
features. IEEE Access 2018;6:31798–807.
[17] Wu Q, Zhu X, Liu B. A survey of android malware static detection technology based on machine learning. Mob Inf Syst 2021;2021:1–18.
[18] Gadyatskaya O, Lezza A-L, Zhauniarovich Y. Evaluation of resource-based app repackaging detection in android. In: Secure IT systems: 21st nordic
conference, nordSec 2016, Oulu, Finland, November 2-4, 2016. Proceedings 21. Springer; 2016, p. 135–51.
[19] Harris MA, Brookshire R, Patten K, Regan B. Mobile application installation influences: have mobile device users become desensitized to excessive
permission requests. In: Proceedings of the twentieth Americas conference on information systems. AMCIS 2015, 2015, p. 13–5.
[20] Alenezi M, Almomani I. Abusing android permissions: A security perspective. In: 2017 IEEE Jordan conference on applied electrical engineering and
computing technologies. AEECT, IEEE; 2017, p. 1–6.
[21] Sharmeen S, Huda S, Abawajy J, Hassan MM. An adaptive framework against android privilege escalation threats using deep learning and semi-supervised
approaches. Appl Soft Comput 2020;89:106089.
[22] Wu D, Cheng Y, Gao D, Li Y, Deng RH. SCLib: A practical and lightweight defense against component hijacking in android applications. In: Proceedings
of the eighth ACM conference on data and application security and privacy. 2018, p. 299–306.
[23] Chin E, Felt AP, Greenwood K, Wagner D. Analyzing inter-application communication in android. In: Proceedings of the 9th international conference on
mobile systems, applications, and services. 2011, p. 239–52.
[24] Choi H, Kim Y. Large-scale analysis of remote code injection attacks in android apps. Secur Commun Netw 2018;2018.
[25] Cai Y, Tang Y, Li H, Yu L, Zhou H, Luo X, He L, Su P. Resource race attacks on android. In: 2020 IEEE 27th international conference on software
analysis, evolution and reengineering. SANER, IEEE; 2020, p. 47–58.
[26] Ruggia A, Possemato A, Merlo A, Nisi D, Aonzo S. Android, notify me when it is time to go phishing. In: 2023 IEEE 8th European symposium on security
and privacy. EuroS&P, IEEE; 2023, p. 1–17.
[27] Ito K, Hasegawa H, Yamaguchi Y, Shimada H. Detecting privacy information abuse by android apps from API call logs. In: Advances in information
and computer security: 13th international workshop on security, IWSEC 2018, Sendai, Japan, September 3-5, 2018, proceedings 13. Springer; 2018, p.
143–57.
[28] Fedler R, Kulicke M, Schütte J. Native code execution control for attack mitigation on android. In: Proceedings of the third ACM workshop on security
and privacy in smartphones & mobile devices. 2013, p. 15–20.
[29] Albanese D, Casolare R, Ciaramella G, Iadarola G, Martinelli F, Mercaldo F, Russodivito M, Santone A, et al. StegWare: A novel malware model exploiting
payload steganography and dynamic compilation. In: ICISSP. 2023, p. 741–8.
[30] Developer. App metadata. 2023, https://developer.apple.com/documentation/appstoreconnectapi/app_store/app_metadata. Accessed on: 2023-12-05.
[31] Developer. Requesting access to protected resources. 2023, https://developer.apple.com/documentation/uikit/protecting_the_user_s_privacy/requesting_
access_to_protected_resources. Accessed on: 2023-12-05.
[32] Localization I. Requesting access to protected resources. 2018, https://medium.com/@guerrix/info-plist-localization-ad5daaea732a. Accessed on:
2023-12-05.

20
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

[33] Inc. G. Platform architecture. 2020, https://developer.android.com/guide/platform. Accessed on: 2023-12-05.

[34] Ng B. Android security overview. 2019, https://medium.com/@boshng95/android-security-overview-7386022ad55d. Accessed on: 2023-12-02.
[35] Khan S, Yusuf A, Haider M, Thirunavukkarasu K, Nand P, Rahmani MKI. A review of android and iOS operating system security. In: 2022 ASU international
conference in emerging technologies for sustainability and intelligent systems. ICETSIS, IEEE; 2022, p. 67–72.
[36] Mohamed I, Patel D. Android vs iOS security: A comparative study. In: 2015 12th international conference on information technology-new generations.
IEEE; 2015, p. 725–30.
[37] Thiel D. iOS application security: The definitive guide for hackers and developers. No Starch Press; 2016.
[38] Security AP. Network security overview. 2021, https://support.apple.com/en-qa/guide/security/sec79afd0274/web. Accessed on: 2023-11-25.
[39] Guo C, Xu J, Liu L, Xu S. MalDetector-using permission combinations to evaluate malicious features of android app. In: 2015 6th IEEE international
conference on software engineering and service science. ICSESS, IEEE; 2015, p. 157–60.
[40] Kaur G, Lashkari AH. Understanding android malware families (UAMF) – the foundations (article 1). 2021, https://www.itworldcanada.com/blog/
understanding-android-malware-families-uamf-the-foundations-article-1/441562. Accessed on: 2023-10-13.
[41] Fiky AHE, Shenawy AE, Madkour MA. Android malware category and family detection and identification using machine learning. 2021, arXiv preprint
arXiv:2107.01927.
[42] Suresh S, Di Troia F, Potika K, Stamp M. An analysis of android adware. J Comput Virol Hack Tech 2019;15(3):147–60.
[43] Ndagi JY, Alhassan JK. Machine learning classification algorithms for adware in android devices: A comparative evaluation and analysis. In: 2019 15th
international conference on electronics, computer and computation. ICECCO, 2019, p. 1–6. http://dx.doi.org/10.1109/ICECCO48375.2019.9043288.
[44] Bagui S, Benson D. Android adware detection using machine learning. Int J Cyber Res Educ (IJCRE) 2021;3(2):1–19.
[45] Daeef AY, Al-Naji A, Chahl J. Features engineering for malware family classification based API call. Computers 2022;11(11):160.
[46] Cho J, Cho G, Hyun S, Kim H. Open sesame! Design and implementation of backdoor to secretly unlock android devices. J Internet Serv Inf Secur
2017;7(4):35–44.
[47] Yao Y, Zhu L, Wang H. Real-time detection of passive backdoor behaviors on android system. In: 2018 IEEE conference on communications and network
security. CNS, 2018, p. 1–9. http://dx.doi.org/10.1109/CNS.2018.8433190.
[48] Li C, Chen X, Wang D, Wen S, Ahmed ME, Camtepe S, Xiang Y. Backdoor attack on machine learning based android malware detectors. IEEE Trans
Dependable Secure Comput 2022;19(5):3357–70. http://dx.doi.org/10.1109/TDSC.2021.3094824.
[49] Zhang Y, Feng F, Liao Z, Li Z, Yao S. Universal backdoor attack on deep neural networks for malware detection. Appl Soft Comput 2023;143:110389.
[50] Kaur G, Lashkari AH. Understanding android malware families: file infector and potentially unwanted applications (article 6). 2021, https:
//www.itworldcanada.com/blog/understanding-android-malware-families-file-infector-and-potentially-unwanted-applications-article-6/455415. Accessed
on: 2023-11-03.
[51] Perlroth N. Android phones hit by ‘ransomware’. 2014, https://archive.nytimes.com/bits.blogs.nytimes.com/2014/08/22/android-phones-hit-by-
ransomware/. Accessed on: 2023-10-10.
[52] News S. Cyberattack has hit at least 74 countries’. 2017, https://news.sky.com/video/cyberattack-has-hit-at-least-74-countries-10874889. Accessed on:
2023-10-04.
[53] Gharib A, Ghorbani A. Dna-droid: A real-time android ransomware detection framework. In: International conference on network and system security.
Springer; 2017, p. 184–98.
[54] Sharma S, Kumar R, Krishna CR. RansomAnalysis: The evolution and investigation of android ransomware. In: Proceedings of international conference
on ioT inclusive life (ICIIL 2019), NITTTR Chandigarh, India. Springer; 2020, p. 33–41.
[55] Amer E, El-Sappagh S. Robust deep learning early alarm prediction model based on the behavioural smell for android malware. Comput Secur
2022;116:102670.
[56] Mahdavifar S, Abdul Kadir AF, Fatemi R, Alhadidi D, Ghorbani AA. Dynamic android malware category classification using semi-supervised deep learning.
In: 2020 IEEE intl conf on dependable, autonomic and secure computing, intl conf on pervasive intelligence and computing, intl conf on cloud and big data
computing, intl conf on cyber science and technology congress. DASC/piCom/cBDCom/cyberSciTech, 2020, p. 515–22. http://dx.doi.org/10.1109/DASC-
PICom-CBDCom-CyberSciTech49142.2020.00094.
[57] Hamid IRA, Khalid NS, Abdullah NA, Ab Rahman NH, Wen CC. Android malware classification using K-means clustering algorithm. In: IOP conference
series: materials science and engineering. Vol. 226, IOP Publishing; 2017, 012105.
[58] Bagui S, Brock H. Machine learning for android scareware detection. J Inf Technol Res (JITR) 2022;15(1):1–15.
[59] Pierazzi F, Mezzour G, Han Q, Colajanni M, Subrahmanian V. A data-driven characterization of modern android spyware. ACM Trans Manage Inf Syst
(TMIS) 2020;11(1):1–38.
[60] Salih HM, Mohammed MS. Spyware injection in android using fake application. In: 2020 international conference on computer science and software
engineering. CSASE, IEEE; 2020, p. 100–5.
[61] Qabalin MK, Naser M, Alkasassbeh M. Android spyware detection using machine learning: A novel dataset. Sensors 2022;22(15):5765.
[62] Abualola H, Alhawai H, Kadadha M, Otrok H, Mourad A. An android-based trojan spyware to study the notificationlistener service vulnerability. Procedia
Comput Sci 2016;83:465–71.
[63] Aminuddin NI, Abdullah Z. Android trojan detection based on dynamic analysis. Adv Comput Intell Syst 2019;1(1):1–7.
[64] Ullah S, Ahmad T, Buriro A, Zara N, Saha S. TrojanDetector: A multi-layer hybrid approach for trojan detection in android applications. Appl Sci
2022;12(21):10755.
[65] Eslahi M, Salleh R, Anuar NB. Mobots: A new generation of botnets on mobile devices and networks. In: 2012 international symposium on computer
applications and industrial electronics. ISCAIE, IEEE; 2012, p. 262–6.
[66] Hashim HA-B, Mohd Saudi M, Basir N. A systematic review analysis of root exploitation for mobile botnet detection. Adv Comput Commun Eng Technol
2016;113–22.
[67] Yerima SY, Alzaylaee MK, Shajan A. Deep learning techniques for android botnet detection. Electronics 2021;10(4):519.
[68] Canadian Institute for Cybersecurity. Android botnet dataset. 2014, https://www.unb.ca/cic/datasets/android-botnet.html. Accessed on: 2023-12-07.
[69] Alani MM, Awad AI. AdStop: Efficient flow-based mobile adware detection using machine learning. Comput Secur 2022;117:102718.
[70] Moreira CC, Moreira DC, de Sales Jr CdS. Improving ransomware detection based on portable executable header using xception convolutional neural
network. Comput Secur 2023;130:103265.
[71] Faghihi F, Zulkernine M. RansomCare: Data-centric detection and mitigation against smartphone crypto-ransomware. Comput Netw 2021;191:108011.
[72] Almohaini R, Almomani I, AlKhayer A. Hybrid-based analysis impact on ransomware detection for android systems. Appl Sci 2021;11(22):10976.
[73] Su D, Liu J, Wang X, Wang W. Detecting android locker-ransomware on chinese social networks. IEEE Access 2018;7:20381–93.
[74] Alsoghyer S, Almomani I. Ransomware detection system for android applications. Electronics 2019;8(8):868.
[75] Faris H, Habib M, Almomani I, Eshtay M, Aljarah I. Optimizing extreme learning machines using chains of salps for efficient android ransomware
detection. Appl Sci 2020;10(11):3706.
[76] Gera T, Singh J, Mehbodniya A, Webber JL, Shabaz M, Thakur D. Dominant feature selection and machine learning-based hybrid approach to analyze
android ransomware. Secur Commun Netw 2021;2021:1–22.
[77] Manavi F, Hamzeh A. Ransomware detection based on PE header using convolutional neural networks.. ISeCure 2022;14(2):181–92.

21
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

[78] Shahzad RK, Lavesson N, Johnson H. Accurate adware detection using opcode sequence extraction. In: 2011 sixth international conference on availability,
reliability and security. IEEE; 2011, p. 189–95.
[79] Fallah S, Bidgoly AJ. Benchmarking machine learning algorithms for android malware detection. Jordanian J Comput Inf Technol 2019;5(3):216–29.
[80] Li G. Research on smartphone trojan detection based on the wireless sensor network. Secur Commun Netw 2022;2022.
[81] Dehkordy DT, Rasoolzadegan A. DroidTKM: Detection of trojan families using the KNN classifier based on manhattan distance metric. In: 2020 10th
international conference on computer and knowledge engineering. ICCKE, IEEE; 2020, p. 136–41.
[82] Hojjatinia S, Hamzenejadi S, Mohseni H. Android botnet detection using convolutional neural networks. In: 2020 28th Iranian conference on electrical
engineering. ICEE, IEEE; 2020, p. 1–6.
[83] Rasheed MM, Faieq AK, Hashim AA. Android botnet detection using machine learning. Ingén Syst Inf 2020;25(1):127–30.
[84] Alqatawna J, Ala’M A-Z, Hassonah MA, Faris H, et al. Android botnet detection using machine learning models based on a comprehensive static analysis
approach. J Inf Secur Appl 2021;58:102735.
[85] Yerima SY, Bashar A. A novel android botnet detection system using image-based and manifest file features. Electronics 2022;11(3):486.
[86] Xue M, Yuan C, Wu H, Zhang Y, Liu W. Machine learning security: Threats, countermeasures, and evaluations. IEEE Access 2020;8:74720–42.
[87] Cinà AE, Grosse K, Demontis A, Biggio B, Roli F, Pelillo M. Machine learning security against data poisoning: Are we there yet? Computer
2024;57(3):26–34.
[88] Yerlikaya FA, Bahtiyar Ş. Data poisoning attacks against machine learning algorithms. Expert Syst Appl 2022;208:118101.
[89] Reith RN, Schneider T, Tkachenko O. Efficiently stealing your machine learning models. In: Proceedings of the 18th ACM workshop on privacy in the
electronic society. 2019, p. 198–210.
[90] Lee T, Edwards B, Molloy I, Su D. Defending against machine learning model stealing attacks using deceptive perturbations. 2018, arXiv preprint
arXiv:1806.00054.
[91] Al-Rubaie M, Chang JM. Privacy-preserving machine learning: Threats and solutions. IEEE Secur Priv 2019;17(2):49–58.
[92] Fredrikson M, Jha S, Ristenpart T. Model inversion attacks that exploit confidence information and basic countermeasures. In: Proceedings of the 22nd
ACM SIGSAC conference on computer and communications security. 2015, p. 1322–33.
[93] Thambiraja E, Ramesh G, Umarani DR. A survey on various most common encryption techniques. Int J Adv Res Comput Sci Softw Eng 2012;2(7):226–33.
[94] Yadav S, Ekbal A, Saha S, Bhattacharyya P. Deep learning architecture for patient data de-identification in clinical records. In: Proceedings of the clinical
natural language processing workshop. clinicalNLP, 2016, p. 32–41.
[95] Löbner S, Tronnier F, Pape S, Rannenberg K. Comparison of de-identification techniques for privacy preserving data analysis in vehicular data sharing.
In: Proceedings of the 5th ACM computer science in cars symposium. 2021, p. 1–11.
[96] Zhou L, Su C, Li Z, Liu Z, Hancke GP. Automatic fine-grained access control in SCADA by machine learning. Future Gener Comput Syst 2019;93:548–59.
[97] Salahdine F, Kaabouch N. Social engineering attacks: A survey. Future Internet 2019;11(4):89.
[98] Hiter S. What is data segmentation? 2021, https://www.datamation.com/security/data-segmentation/. Accessed on: 2023-10-16.
[99] Baracaldo N, Chen B, Ludwig H, Safavi JA. Mitigating poisoning attacks on machine learning models: A data provenance based approach. In: Proceedings
of the 10th ACM workshop on artificial intelligence and security. 2017, p. 103–10.
[100] Xu R, Baracaldo N, Joshi J. Privacy-preserving machine learning: Methods, challenges and directions. 2021, arXiv preprint arXiv:2108.04417.
[101] Sun X, Zhang P, Liu JK, Yu J, Xie W. Private machine learning classification based on fully homomorphic encryption. IEEE Trans Emerg Top Comput
2018;8(2):352–64.
[102] Vijayarani S, Tamilarasi A, Sampoorna M. Analysis of privacy preserving k-anonymity methods and techniques. In: 2010 international conference on
communication and computational intelligence. INCOCCI, IEEE; 2010, p. 540–5.
[103] Friedman A, Wolff R, Schuster A. Providing k-anonymity in data mining. VLDB J 2008;17:789–804.
[104] Kim W, Seok J. Privacy-preserving collaborative machine learning in biomedical applications. In: 2022 international conference on artificial intelligence
in information and communication. ICAIIC, IEEE; 2022, p. 179–83.
[105] Zhao C, Zhao S, Zhao M, Chen Z, Gao C-Z, Li H, Tan Y-a. Secure multi-party computation: theory, practice and applications. Inform Sci 2019;476:357–72.
[106] Amro B. Malware detection techniques for mobile devices. 2018, arXiv preprint arXiv:1801.02837.
[107] Milosevic N, Dehghantanha A, Choo K-KR. Machine learning aided android malware classification. Comput Electr Eng 2017;61:266–74.
[108] Arshad S, Shah MA, Wahid A, Mehmood A, Song H, Yu H. SAMADroid: A novel 3-level hybrid malware detection model for android operating system.
IEEE Access 2018;6:4321–39. http://dx.doi.org/10.1109/ACCESS.2018.2792941.
[109] Syrris V, Geneiatakis D. On machine learning effectiveness for malware detection in android os using static analysis data. J Inf Secur Appl 2021;59:102794.
[110] Suarez-Tangil G, Dash SK, Ahmadi M, Kinder J, Giacinto G, Cavallaro L. DroidSieve: Fast and accurate classification of obfuscated android malware.
CODASPY ’17, New York, NY, USA: Association for Computing Machinery; 2017, p. 309–20. http://dx.doi.org/10.1145/3029806.3029825.
[111] Mat SRT, Ab Razak MF, Kahar MNM, Arif JM, Firdaus A. A Bayesian probability model for android malware detection. ICT Express 2022;8(3):424–31.
[112] Bayazit EC, Sahingoz OK, Dogan B. A deep learning based android malware detection system with static analysis. In: 2022 international congress on
human-computer interaction, optimization and robotic applications. HORA, IEEE; 2022, p. 1–6.
[113] Yang H, Wang Y, Zhang L, Cheng X, Hu Z. A novel android malware detection method with API semantics extraction. Comput Secur 2024;137:103651.
[114] HR S. Static analysis of android malware detection using deep learning. In: 2019 international conference on intelligent computing and control systems.
ICCS, 2019, p. 841–5. http://dx.doi.org/10.1109/ICCS45141.2019.9065765.
[115] Singh AK, Jaidhar C, Kumara M. Experimental analysis of android malware detection based on combinations of permissions and API-calls. J Comput
Virol Hack Tech 2019;15(3):209–18.
[116] AndroZoo. AndroZoo. 2023, https://androzoo.uni.lu/. Accessed on: 2023-12-24.
[117] DREBIN. The drebin dataset. 2023, https://www.sec.tu-bs.de/~danarp/drebin/download.html. Accessed on: 2023-12-24.
[118] Akbar F, Hussain M, Mumtaz R, Riaz Q, Wahab AWA, Jung K-H. Permissions-based detection of android malware using machine learning. Symmetry
2022;14(4):718.
[119] VIRUSSHARE. VIRUSSHARE. 2023, https://www.secrepo.com/Datasets%20Description/PE_malware/VirusShare.html. Accessed on: 2023-12-24.
[120] İbrahim M, Issa B, Jasser MB. A method for automatic android malware detection based on static analysis and deep learning. IEEE Access
2022;10:117334–52.
[121] Mahdavifar S, Alhadidi D, Ghorbani AA. Effective and efficient hybrid android malware classification using pseudo-label stacked auto-encoder. J Netw
Syst Manage 2022;30:1–34.
[122] Odat E, Yaseen QM. A novel machine learning approach for android malware detection based on the co-existence of features. IEEE Access
2023;11:15471–84.
[123] Chaudhary M, Masood A. RealMalSol: real-time optimized model for android malware detection using efficient neural networks and model quantization.
Neural Comput Appl 2023;35(15):11373–88.
[124] Ding Y, Zhang X, Hu J, Xu W. Android malware detection method based on bytecode image. J Ambient Intell Humaniz Comput 2023;14(5):6401–10.
[125] Nguyen C-D, Khoa NH, Doan KN-D, Cam NT. Android malware category and family classification using static analysis. In: 2023 international conference
on information networking. ICOIN, IEEE; 2023, p. 162–7.
[126] Ariffin NAM, Casinto HP. Android malware detection using permission based static analysis. J Adv Res Appl Sci Eng Technol 2023;33(3):86–97.

22
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

[127] Bakır H, Bakır R. DroidEncoder: Malware detection using auto-encoder based feature extractor and machine learning algorithms. Comput Electr Eng
2023;110:108804.
[128] Garcia J, Hammad M, Pedrood B, Bagheri-Khaligh A, Malek S. Obfuscation-resilient, efficient, and accurate detection and family identification of android
malware. Tech. rep., 202, Department of Computer Science, George Mason University; 2015.
[129] Zhang X, Breitinger F, Luechinger E, O’Shaughnessy S. Android application forensics: A survey of obfuscation, obfuscation detection and deobfuscation
techniques and their impact on investigations. Forensic Sci Int: Digit Invest 2021;39:301285.
[130] Poeplau S, Fratantonio Y, Bianchi A, Kruegel C, Vigna G. Execute this! analyzing unsafe and malicious dynamic code loading in android applications.
In: NDSS. Vol. 14, 2014, p. 23–6.
[131] Yan P, Yan Z. A survey on dynamic mobile malware detection. Softw Qual J 2018;26(3):891–919.
[132] Haq IU, Khan TA, Akhunzada A. A dynamic robust DL-based model for android malware detection. IEEE Access 2021;9:74510–21. http://dx.doi.org/10.
1109/ACCESS.2021.3079370.
[133] Sihag V, Vardhan M, Singh P, Choudhary G, Son S. De-LADY: Deep learning based android malware detection using dynamic features. J Internet Serv
Inf Secur 2021;11(2):34–45.
[134] Bhat P, Behal S, Dutta K. A system call-based android malware detection approach with homogeneous & heterogeneous ensemble machine learning.
Comput Secur 2023;130:103277.
[135] Liu Z, Wang R, Japkowicz N, Gomes HM, Peng B, Zhang W. SeGDroid: An android malware detection method based on sensitive function call graph
learning. Expert Syst Appl 2024;235:121125.
[136] Mahindru A, Singh P. Dynamic permissions based android malware detection using machine learning techniques. In: Proceedings of the 10th innovations
in software engineering conference. ISEC ’17, New York, NY, USA: Association for Computing Machinery; 2017, p. 202–10. http://dx.doi.org/10.1145/
3021460.3021485.
[137] Abdul Kadir AF, Stakhanova N, Ghorbani AA. Android botnets: What urls are telling us. In: Network and system security: 9th international conference,
NSS 2015, New York, NY, USA, November 3-5, 2015, proceedings 9. Springer; 2015, p. 78–91.
[138] Gonzalez H, Stakhanova N, Ghorbani AA. Droidkin: Lightweight detection of android apps similarity. In: International conference on security and privacy
in communication networks: 10th international ICST conference, secureComm 2014, Beijing, China, September 24-26, 2014, revised selected papers, part
i 10. Springer; 2015, p. 436–53.
[139] Feng P, Ma J, Sun C, Xu X, Ma Y. A novel dynamic android malware detection system with ensemble learning. IEEE Access 2018;6:30996–1011.
http://dx.doi.org/10.1109/ACCESS.2018.2844349.
[140] Li C, Lv Q, Li N, Wang Y, Sun D, Qiao Y. A novel deep framework for dynamic malware detection based on API sequence intrinsic features. Comput
Secur 2022;116:102686.
[141] Wang L, Wang H, He R, Tao R, Meng G, Luo X, Liu X. MalRadar: Demystifying android malware in the new era. Proc ACM Meas Anal Comput Syst
2022;6(2):1–27.
[142] Chaganti R, Ravi V, Pham TD. A multi-view feature fusion approach for effective malware classification using deep learning. J Inf Secur Appl
2023;72:103402.
[143] Wang H, Zhang W, He H. You are what the permissions told me! android malware detection based on hybrid tactics. J Inf Secur Appl 2022;66:103159.
[144] Samaneh M, Dima A, Ghorbani AA. Effective and efficient hybrid android malware classification using pseudo-label stacked auto-encoder. J Netw Syst
Manage 2022;30(1):22.
[145] Taher F, AlFandi O, Al-kfairy M, Al Hamadi H, Alrabaee S. DroidDetectMW: A hybrid intelligent model for android malware detection. Appl Sci
2023;13(13):7720.
[146] Dhalaria M, Gandotra E. MalDetect: A classifier fusion approach for detection of android malware. Expert Syst Appl 2024;235:121155.
[147] Surendran R, Thomas T, Emmanuel S. A TAN based hybrid model for android malware detection. J Inf Secur Appl 2020;54:102483.
[148] Hadiprakoso RB, Kabetta H, Buana IKS. Hybrid-based malware analysis for effective and efficiency android malware detection. In: 2020 international
conference on informatics, multimedia, cyber and information system. ICIMCIS, IEEE; 2020, p. 8–12.
[149] Lu T, Du Y, Ouyang L, Chen Q, Wang X. Android malware detection based on a hybrid deep learning model. Secur Commun Netw 2020;2020:1–11.
[150] Dhalaria M, Gandotra E. A hybrid approach for android malware detection and family classification. 2020, p. 174–88.
[151] Upadhayay M, Sharma A, Garg G, Arora A. Rpndroid: Android malware detection using ranked permissions and network traffic. In: 2021 fifth world
conference on smart trends in systems security and sustainability. WorldS4, IEEE; 2021, p. 19–24.
[152] KOODOUS. KOODOUS. 2023, https://koodous.com/. Accessed on: 2023-12-11.
[153] Lashkari AH, Kadir AFA, Taheri L, Ghorbani AA. Toward developing a systematic approach to generate benchmark android malware datasets and
classification. In: 2018 international carnahan conference on security technology. ICCST, IEEE; 2018, p. 1–7.
[154] Asad M, Moustafa A, Aslam M. CEEP-FL: A comprehensive approach for communication efficiency and enhanced privacy in federated learning. Appl Soft
Comput 2021;104:107235.
[155] Li L, Fan Y, Tse M, Lin K-Y. A review of applications in federated learning. Comput Ind Eng 2020;149:106854.
[156] Li T, Sahu AK, Talwalkar A, Smith V. Federated learning: Challenges, methods, and future directions. IEEE Signal Process Mag 2020;37(3):50–60.
[157] Rieke N, Hancox J, Li W, Milletari F, Roth HR, Albarqouni S, Bakas S, Galtier MN, Landman BA, Maier-Hein K, et al. The future of digital health with
federated learning. NPJ Digit Med 2020;3(1):119.
[158] Khan M, Glavin FG, Nickles M. Federated learning as a privacy solution - an overview. Procedia Comput Sci 2023;217:316–25. http://dx.doi.org/10.
1016/j.procs.2022.12.227, URL https://www.sciencedirect.com/science/article/pii/S1877050922023055. 4th International Conference on Industry 4.0 and
Smart Manufacturing.
[159] Lin BY, He C, Zeng Z, Wang H, Huang Y, Soltanolkotabi M, Ren X, Avestimehr S. Fednlp: A research platform for federated learning in natural language
processing. 2021, arXiv preprint arXiv:2104.08815.
[160] Toldinas J, Venčkauskas A, Liutkevičius A, Morkevičius N. Framing network flow for anomaly detection using image recognition and federated learning.
Electronics 2022;11(19):3138.
[161] Xu J, Glicksberg BS, Su C, Walker P, Bian J, Wang F. Federated learning for healthcare informatics. J Healthc Inform Res 2021;5:1–19.
[162] Zhang T, Gao L, He C, Zhang M, Krishnamachari B, Avestimehr AS. Federated learning for the internet of things: applications, challenges, and opportunities.
IEEE Internet Things Mag 2022;5(1):24–9.
[163] Nguyen TD, Rieger P, Miettinen M, Sadeghi A-R. Poisoning attacks on federated learning-based IoT intrusion detection system. In: Proc. workshop
decentralized IoT syst. secur.. 2020, p. 1–7.
[164] Yu B, Mao W, Lv Y, Zhang C, Xie Y. A survey on federated learning in data mining. Wiley Interdiscip Rev: Data Min Knowl Discov 2022;12(1):e1443.
[165] Qayyum A, Ahmad K, Ahsan M, Al-Fuqaha A, Qadir J. Collaborative federated learning for healthcare: Multi-modal COVID-19 diagnosis at the edge.
IEEE Open J Comput Soc 2022;3(01):172–84. http://dx.doi.org/10.1109/OJCS.2022.3206407.
[166] Alamer A. A privacy-preserving federated learning with a secure collaborative for malware detection models using internet of things resources. Internet
Things 2024;25:101015.
[167] Tolpegin V, Truex S, Gursoy ME, Liu L. Data poisoning attacks against federated learning systems. In: Computer security–ESORICs 2020: 25th European
symposium on research in computer security, ESORICs 2020, Guildford, UK, September 14–18, 2020, proceedings, part i 25. Springer; 2020, p. 480–501.
[168] Zhou X, Xu M, Wu Y, Zheng N. Deep model poisoning attack on federated learning. Future Internet 2021;13(3):73.

23
F. Nawshin et al. Computers and Electrical Engineering 117 (2024) 109233

[169] Bhagoji AN, Chakraborty S, Mittal P, Calo S. Analyzing federated learning through an adversarial lens. In: International conference on machine learning.
PMLR; 2019, p. 634–43.
[170] Li L, Fan Y, Tse M, Lin K-Y. A review of applications in federated learning. Comput Ind Eng 2020;149:106854.
[171] Ding J, Tramel E, Sahu AK, Wu S, Avestimehr S, Zhang T. Federated learning challenges and opportunities: An outlook. In: ICASSP 2022-2022 IEEE
international conference on acoustics, speech and signal processing. ICASSP, IEEE; 2022, p. 8752–6.
[172] Cao X, Sun G, Yu H, Guizani M. PerFED-GAN: Personalized federated learning via generative adversarial networks. IEEE Internet Things J 2022;3749–62.
[173] Wang X, Zhu T, Zhou W. Supplement data in federated learning with a generator transparent to clients. Inform Sci 2024;120437. http://dx.doi.org/10.
1016/j.ins.2024.120437.
[174] Li Z, Duan M, Yu S, Yang W. DynamicNet: Efficient federated learning for mobile edge computing with dynamic privacy budget and aggregation weights.
IEEE Trans Consum Electron 2024. http://dx.doi.org/10.1109/TCE.2024.3372696, 1–1.
[175] Chen X, Du T, Wang M, Gu T, Zhao Y, Kou G, Xu C, Wu. DO. Towards optimal customized architecture for heterogeneous federated learning with
contrastive cloud-edge model decoupling. 2024, arXiv preprint arXiv:2403.02360.
[176] Azimi-Abarghouyi, Mohammad S, Fodor V. Quantized hierarchical federated learning: A robust approach to statistical heterogeneity. 2024, arXiv preprint
arXiv:2403.01540.
[177] Gálvez R, Moonsamy V, Diaz C. Less is more: A privacy-respecting android malware classifier using federated learning. 2020, arXiv preprint arXiv:
2007.08319.
[178] Jiang C, Yin K, Xia C, Huang W. FedHGCDroid: An adaptive multi-dimensional federated learning for privacy-preserving android malware classification.
Entropy 2022;24(7):919.
[179] Mahindru A, Arora H. Dnndroid: Android malware detection framework based on federated learning and edge computing. In: Advancements in smart
computing and information security: first international conference, ASCIS 2022, Rajkot, India, November 24–26, 2022, revised selected papers, part II.
Springer; 2023, p. 96–107.
[180] Chaudhuri A, Nandi A, Pradhan B. A dynamic weighted federated learning for android malware classification. In: Soft computing: theories and applications:
proceedings of soCTA 2022. Springer; 2023, p. 147–59.
[181] Rey V, Sánchez PMS, Celdrán AH, Bovet G. Federated learning for malware detection in iot devices. Comput Netw 2022;204:108693.
[182] Xia Q, Ye W, Tao Z, Wu J, Li Q. A survey of federated learning for edge computing: Research problems and solutions. High-Confid Comput
2021;1(1):100008.
[183] Hsu R-H, Wang Y-C, Fan C-I, Sun B, Ban T, Takahashi T, Wu T-W, Kao S-W. A privacy-preserving federated learning system for android malware
detection based on edge computing. In: 2020 15th Asia joint conference on information security. AsiajCIS, 2020, p. 128–36. http://dx.doi.org/10.1109/
AsiaJCIS50894.2020.00031.
[184] Guerra-Manzanares A, Bahsi H, Nõmm S. Kronodroid: Time-based hybrid-featured dataset for effective android malware detection and characterization.
Comput Secur 2021;110:102399.
[185] TUANDROMD. TUANDROMD (tezpur university android malware dataset). 2023, https://www.archive.ics.uci.edu/dataset/855/tuandromd+(tezpur+
university+android+malware+dataset). Accessed on: 2023-12-20.
[186] Amjath M, Henna S. Differential Privacy Preservation for Graph-based Federated Learning under Malware Attacks.
[187] Jin X, Chen P-Y, Hsu C-Y, Yu C-M, Chen T. CAFE: Catastrophic data leakage in vertical federated learning. Adv Neural Inf Process Syst 2021;34:994–1006.
[188] Benmalek M, Benrekia MA, Challal Y. Security of federated learning: attacks, defensive mechanisms, and challenges. Rev Sci Technol l’Inf-Sér RIA: Rev
d’Intell Artif 2022;36(1):49–59.
[189] Shi J, Wan W, Hu S, Lu J, Zhang LY. Challenges and approaches for mitigating byzantine attacks in federated learning. In: 2022 IEEE international
conference on trust, security and privacy in computing and communications. TrustCom, IEEE; 2022, p. 139–46.
[190] Chen Y, Gui Y, Lin H, Gan W, Wu Y. Federated learning attacks and defenses: A survey. 2022, arXiv preprint arXiv:2211.14952.
[191] Bhagoji AN, Chakraborty S, Mittal P, Calo S. Analyzing federated learning through an adversarial lens. In: International conference on machine learning.
PMLR; 2019, p. 634–43.
[192] Doku R, Rawat DB. Mitigating data poisoning attacks on a federated learning-edge computing network. In: 2021 IEEE 18th annual consumer
communications & networking conference. CCNC, IEEE; 2021, p. 1–6.
[193] Psychogyios K, Velivassaki T-H, Bourou S, Voulkidis A, Skias D, Zahariadis T. GAN-driven data poisoning attacks and their mitigation in federated learning
systems. Electronics 2023;12(8):1805.
[194] Rakin AS, He Z, Fan D. Tbt: Targeted neural network attack with bit trojan. In: Proceedings of the IEEE/CVF conference on computer vision and pattern
recognition. 2020, p. 13198–207.
[195] Zhang J, Zhu H, Wang F, Zhao J, Xu Q, Li H, et al. Security and privacy threats to federated learning: Issues, methods, and challenges. Secur Commun
Netw 2022;2022.
[196] Chai H, Leng S, Chen Y, Zhang K. A hierarchical blockchain-enabled federated learning algorithm for knowledge sharing in internet of vehicles. IEEE
Trans Intell Transp Syst 2020;22(7):3975–86.
[197] Uprety A, Rawat DB. Mitigating poisoning attack in federated learning. In: 2021 IEEE symposium series on computational intelligence. SSCI, IEEE; 2021,
p. 01–7.
[198] Unal D, Hammoudeh M, Khan MA, Abuarqoub A, Epiphaniou G, Hamila R. Integration of federated machine learning and blockchain for the provision
of secure big data analytics for internet of things. Comput Secur 2021;109:102393. http://dx.doi.org/10.1016/j.cose.2021.102393.
[199] Ali M, Karimipour H, Tariq M. Integration of blockchain and federated learning for internet of things: Recent advances and future challenges. Comput
Secur 2021;108(C):102355. http://dx.doi.org/10.1016/j.cose.2021.102355.
[200] Abou El Houda Z, Hafid AS, Khoukhi L. MiTFed: A privacy preserving collaborative network attack mitigation framework based on federated learning
using SDN and blockchain. IEEE Trans Netw Sci Eng 2023;1985–2001.
[201] Liu W, He Y, Wang X, Duan Z, Liang W, Liu Y. BFG: privacy protection framework for internet of medical things based on blockchain and federated
learning. Connect Sci 2023;35(1):2199951.

24