Software Design Level Vulnerability Classification Model

Shabana Rehman & Khurram Mustafa
Software Design Level Vulnerability Classification Model

Shabana Rehman
shabana.infosec@gmail.com
Department of Information System

Salman bin Abdul Aziz University
Al-Kharj, KSA
Khurram Mustafa
kmustafa@jmi.ac.in
Department of Computer Science

Jamia Millia Islamia
New Delhi, 110025, India
Abstract
Classification of software security vulnerability no doubt facilitates the understanding of securityrelated information and accelerates vulnerability analysis. The lack of proper classification not
only hinders its understanding but also renders the strategy of developing mitigation mechanism
for clustered vulnerabilities. Now software developers and researchers are agreed on the fact that
requirement and design phase of the software are the phases where security incorporation yields
maximum benefits. In this paper we have attempted to design a classifier that can identify and
classify design level vulnerabilities. In this classifier, first vulnerability classes are identified on the
basis of well established security properties like authentication and authorization. Vulnerability
training data is collected from various authentic sources like Common Weakness Enumeration
(CWE), Common Vulnerabilities and Exposures (CVE) etc. From these databases only those
vulnerabilities were included whose mitigation is possible at the design phase. Then this
vulnerability data is pre-processed using various processes like text stemming, stop word
removal, cases transformation. After pre-processing, SVM (Support Vector Machine) is used to
classify vulnerabilities. Bootstrap validation is used to test and validate the classification process
performed by the classifier. After training the classifier, a case study is conducted on NVD
(National Vulnerability Database) design level vulnerabilities. Vulnerability analysis is done on the
basis of classification result.
Keywords: Security Vulnerabilities, Design Phase, Classification, Machine Leaning, Security Properties
1. INTRODUCTION
Developing secure software remains a significant challenge for todays software developers as
they still face difficulty in understanding the reasons of vulnerabilities in the existing software. It is
vital to be able to identify software security vulnerabilities in the early phases of SDLC (Software
Development Lifecycle) and one of early detection approaches is to consult with the prior known
vulnerabilities and corresponding fixes [1]. Identification of candidate security vulnerability pays a
substantial benefit when they are deals in early phases like requirement and design phases of the
software [2]. Classification of vulnerabilities is fruitful in understanding the vulnerabilities better
and classification also helps in mitigating group of vulnerabilities. Identifying and mitigating
security vulnerabilities is no doubt a difficult task therefore taxonomy is developed that can
classify vulnerabilities into classes and this will help designer to mitigate cluster of vulnerabilities.
There are number of approaches of taxonomy development in past, like [3,4,5,6] etc, but no one
ever propose any taxonomy that classify design level vulnerabilities on the basis of security
properties. We have already proposed a taxonomy in [7], as shown in Table 1.0 (a) in which,
priori classification is proposed and vulnerabilities are classified manually. But in this classification
there is chance of Hawthorne Effect, it also largely depends on the expertise of the classifier.
Therefore here we are creating a classifier that can classify a vulnerability data automatically.
Machine learning is now a popular tool in the automation task. Researchers have explored the
International Journal of Computer Science and Security (IJCSS), Volume (6): Issue (4)
238
use of machine learning techniques to automatically associate documents with categories by first
using a training set to adapt the classifier to the feature set of the particular document set [8].
Machine learning is a relatively new approach that can be used in classifying vulnerabilities.
Therefore here Classifier is proposed, that classifies vulnerabilities on the basis of previously
identified vulnerability and can help designer to place vulnerability in the predefined vulnerability
classes that are based on the security properties of the software. Therefore mitigation mechanism
can be applied for the whole class of vulnerabilities. In this classifier, first data pre-processing is
done like text stemming, stop word removal, case transformation then SVM (Support Vector
Machine) is used for the final classification with the regression model. Several conclusions are
drawn after applying a classification. At last using this classifier NVD (National Vulnerability
Database) vulnerabilities are classified and analyzed.
First Level
Second Level
Third level
Access Control
at Process
Level
Authentication
Authorization
Audit & logging
Access
Control
Access Control
at
Communication
Level
Secured Session
Management
Secured
Information Flow
Exposures
leading to
Access
Violation
Exposures in
Error Message
Predictable
Algorithm
/sequence
numbers/file
names
User Alertness
Fourth Level
Missing Authentication procedure
Insufficient Authentication procedure
Wrong Authentication procedure
Missing Authorization procedure
Insufficient Authorization procedure
Wrong Authorization procedure
Missing Audit and logging
Insufficient Logging or Audit of
information
Wrong Audit or Logging of information
Missing
Secured
Session
management
Insufficient
Secured
Session
Management
Wrong Secured Session Management
Missing Encryption of Sensitive Data
During Transmission
Insufficient Encryption of Sensitive
Data during Transmission
Wrong Encryption of Sensitive Data
during Transmission
Missing Secured Error Message
Insufficient Secured Error Message
Wrong Secured Error Message
Missing Randomness in the Random
Sequence Ids
Insufficient Randomness
in the
Random Sequence Ids
Wrong Randomness in the Random
Sequence Ids or Wrong Choice of File
Name
Missing User Alerting Information
Insufficient User Alerting Information
Wrong User Alerting Information
TABLE 1.0 (a): Taxonomy of Design Level Vulnerabilities
While considering the number of classes in the proposed classifier, we consider only access
control at process level and access control at communication level and all the other type of
vulnerabilities are considered in the Others class. Because exposure leading to access violation
239
class covers a large domain of vulnerabilities and needs a separate study, therefore after
exploring the domain of this class, we exclude this from the classifier and will consider for the
future work.
Rest of the paper is organized as follows, in section 2, related works in the vulnerability
classification is discussed, and then in section 3, the development process of vulnerabilities
classification model is explained in detail. Classification of vulnerabilities using developed
classifier is done in section 4. Conclusion and future work are discussed in section 5.
2. RELATED WORK
There are many classification approaches using machine learning techniques like [9] proposed
uses a ontological approach to retrieving vulnerability data and establishing a relationship
between them, they also reason about the cause and impact of vulnerabilities. In their ontology
vulnerability management (OVM), they have populated all vulnerabilities of NVD (National
Vulnerability Database), with additional inference rules, knowledge representation, and datamining mechanisms. Another relevant work in vulnerability classification area is done by [10], they
proposed a CVE categorization framework that transforms the vulnerability dictionary into a
classifier that categorizes CVE( Common Vulnerability and Exposure) with respect to diverse
taxonomic features and evaluates general trends in the evolution of vulnerabilities. [11], in their
paper, entitled Secure software Design in Practice presented a SODA (a Security-Oriented
Software Development Framework), which was the result of a research project where the main
goal had been to create a system of practical techniques and tools for creating secure software
with a special focus on the design phase of the software. Another approach of categorizing
vulnerabilities is that of [12]. In their paper [12], they looked at the possibilities of categorizing
vulnerabilities in the CVE using SOM. They presented a way to categorize the vulnerabilities in
the CVE repository and proposed a solution for standardization of the vulnerability categories
using a data-clustering algorithm. [13], proposed SecureSync, an automatic approach to detect
and provide suggested resolutions for recurring software vulnerabilities on multiple systems
sharing/using similar code or API libraries. There are many other vulnerability classification
approaches like [14,15,16], but all the above mentioned approaches are either to generic in
nature or they cannot be used to classify vulnerabilities on the basis of security properties of the
software. Therefore in this research work we are proposing a classifier that is developed using
machine learning techniques and is very specific to the design phase of the software. In the next
section, a development stage of the classifier is explained.
3.0 DESIGN LEVEL VULNERABILITIES CLASSIFICATION MODEL

Software vulnerability databases are essential part of software security knowledgebase. There
are a number of vulnerability databases that exchanges software vulnerability information
however, its impact on vulnerability analysis is hindered by its lack of categorization and
generalization functionalities [10]. To extract useful and relevant information from these
databases, lots of manual work is required. For example, if software developer wants to know
about the most common and severe vulnerability prevailing in a current software in a particular
period of time then he has to study all the vulnerability descriptions published during that period,
then he has to classify those vulnerability based on his own criteria and then he has to check the
severity rating provided by various experts. This is very unreliable, tedious and protracted task.
Using a proposed classification model, researchers and developers can easily classify design
level vulnerabilities and identify a mitigation mechanism in the form of design pattern, in the early
phases of the SDLC. The classification results with severity rating can further be used to calculate
the risk of vulnerability occurrence at the design phase of the software.
Automated text classifier is basically used to classify the text in the predefined classes. An
abstract view of the classifier is shown in Fig 3.0 (a). In proposed design level vulnerability
classifier, first text is pre-processed using various processes like tokenization, case
transformation, stop-word removal and stemming, then SVM (Support Vector Machine) is used to
classify the text and finally bootstrap validation is used to test and validate the results. The
development process of the classifier is explained in Fig 3.0 (b).
240
FIGURE 3.0 (a): Abstract Vulnerability Classifier
The vulnerability categorization framework proposed by [10] is similar to this design level
classifier. But Chens framework is a generalized categorization framework that is developed to
classify all the vulnerabilities of CVE, on the bases of classification categories of BID, X-force and
Secunia. The training data in their framework is also taken from these vulnerabilities databases
only.
FIGURE 3.0 (b): Design Level Vulnerabilities Classification Process
In our design level vulnerability classifier, only design level vulnerabilities are classified and in
training data only those identified vulnerabilities are considered which can be mitigated at the
design level of the software. Moreover the classes are defined on the basis of security properties
of the software like authentication, authorization etc., which are generally considered while
developing the security design patterns of the software. Therefore after classification
developers/researchers can priorities prevailing vulnerabilities class before choosing security
design pattern.
3.1 Feature Vector Creation
The vulnerabilities in the CVE are defined in the natural language form. Therefore only way to
identify a feature vector using the vulnerability description is the frequency of keywords in the
description. Therefore feature vector are identified by the keywords used in the description of the
vulnerabilities. To make vulnerability description into a structured representation that can be used
by machine learning algorithms, first the text will be converted into tokens and after stemming and
241
stop word removal, and case transformation. There are five steps in the feature creation process,
specified as follows:
a). Tokenization
b). Case Transformation
c). Stopword elimination
d). Text stemming of CVE entries
e). Weight Assignment
a).Tokenization
The isolation of word-like units form a text is called tokenization. It is a process in which text
stream is to break down into words, phrases and symbols called tokens [17]. These tokens can
be further used as input for the information processing. In order to convert text in machine
learning form, first the raw text is transformed into a machine readable form, and first step
towards it is a tokenization. As shown in Fig. 3.1 (a), the raw text is first feed to the pre-processor,
convert the text in the form of tokens then further morphological analysers are used to perform
required linguistic analysis.
FIGURE 3.1 (a): Text transformations before linguistic analysis
In order to feed vulnerability description in machine learning process, the textual description of
vulnerability is first converted in the form of tokens. In Table 3.1 (a), a vulnerabilities description is
shown after tokenization.
Vulnerability ID
Vulnerability Description
CVE-2007-0164
Camouflage 1.2.1 embeds password

information in the carrier file, which allows
remote attackers to bypass authentication
requirements and decrypt embedded
steganography by replacing certain bytes
of the JPEG image with alternate
password information.
Vulnerability Description after

Tokenization
Camouflage, embeds, password,
information, in, the, carrier, file,
which, allows, remote, attackers,
to,
bypass,
authentication,
requirements,
and,
decrypt,
embedded, steganography, by,
replacing, certain, bytes, of, the,
JPEG, image, with, alternate,
password, information.
TABLE 3.1 (a): Tokenization of Vulnerabilities
242
b). Case Transformation

When raw text is retrieved for processing from any source, then it contains words in both upper
case as well as lower case. The machine learning algorithms reads words in different cases as
different words. In order to transform all the words in the same case, Case transformation
process is used. Case transformer, transforms all characters in a document to either lower case
or upper case, respectively. In our case we have transformed all the words in the document in
lower case.
Vulnerability
ID
CVE-20070164
Camouflage,
embeds,
password,
information, in, the, carrier, file, which,
allows, remote, attackers, to, bypass,
authentication,
requirements,
and,
decrypt, embedded, steganography, by,
replacing, certain, bytes, of, the, JPEG,
image,
with,
alternate,
password,
information

Case Transformation
camouflage, embeds, password,
information, in, the, carrier, file,
which, allows, remote, attackers,
to,
bypass,
authentication,
requirements,
and,
decrypt,
embedded, steganography, by,
replacing, certain, bytes, of, the,
jpge, image, with, alternate,
password, information
TABLE 3.1 (b): Case Transformation of Vulnerabilities Description
c). Stop word Removal

Stop word elimination is a process of removing those tokens that are considered as only for
grammatical function without adding new meaning to sentences they involve [18]. The stop word
list generally consists of articles, case particles, conjunctions, pronouns, auxiliary verbs and
common prepositions. There is no unique list of stop words which is always used. There are
number of lists that are proposed by different researchers. A list of 418 stop word is used by
Chen [10]. A similar stop word list is used in information retrieval systems Snowball [19] and
Lemur [20].The vulnerability description after stop word removal is shown in Table 3.1 (c)
Vulnerability ID
CVE-2007-0164
Vulnerability
description
after
tokenization and case transformation
camouflage,
embeds,
password,
information, in, the, carrier, file, which,
allows, remote, attackers, to, bypass,
authentication,
requirements,
and,
decrypt, embedded, steganography, by,
replacing, certain, bytes, of, the, jpge,
image, with, alternate, password,
information.
Vulnerability
description
after
Stopword Removal
camouflage, embeds, password
,information, carrier, file, allows,
remote,
attackers,
bypass,
authentication,
requirements,
decrypt,
embedded,
steganography, replacing, bytes,
jpge, image, alternate, password,
information.
TABLE 3.1 (c): Stop word removal from vulnerabilities
d). Text Stemming

Uses of stemming algorithms in modern information retrieval (IR) systems are common these
days. Stemming algorithms are helpful for free text retrieval, where search terms can occur in
various different forms in the document collection. Stemming makes retrieval of such documents
independent from the specific word form used in the query [21]. To extract the information from
vulnerability description, stemming algorithm can be used, so that text can be easily transformed
into a machine readable form. Porter stemming algorithm is one of the popular algorithms that is
generally used in the information retrieval process. Porter's algorithm consists of 5 phases of
word reductions, applied sequentially. Within each phase there are various conventions to select
rules, such as selecting the rule from each rule group that applies to the longest suffix. In the first
phase, this convention is used with the following rule group [22]:
Rule
Example
243
SSES
SS
caresses
caress
IES
I
ponies
poni
SS
SS
caress
caress
S
cats
cat
The vulnerability description after applying porter stemming algorithm is shown in Table 3.1 (d).
Vulnerability ID
CVE-2007-0164
Camouflage 1.2.1 embeds password

information in the carrier file, which allows
remote attackers to bypass authentication
requirements and decrypt embedded
steganography by replacing certain bytes
of the JPEG image with alternate
password information.

Text Stemming
camouflag emb password inform
carrier file allow remot attack
bypass authent requir decrypt
embed steganographi replac byte
jpeg imag altern password inform
TABLE 3.1 (d): Stemming of words in Vulnerabilities Description
e). Weight Assignment

After text stemming, the next step is a weight assignment of each word of vulnerability
description. The simplest approach is to assign the weight to be equal to the number of
occurrences of term t in document d. This weighting scheme is referred to as term frequency
and is denoted tf t,d with the subscripts denoting the term and the document in order [22]. It is
normally computed as follows.
tft,d
ft,d / maxk fk,d
Eq. 3.1 (e.1)
where
ft,d is the frequency (number of occurrence of the term t in document d ) and
maxk f k,d (maximum number of occurrences of any term)
Thus, the most frequent term in document d gets a TF as 1, and other terms get fractions as
their term frequency for this document. But the disadvantage of using a term frequency is that, in
this all the terms are considered equally important. In order to avoid this bias, term frequency and
inverse document frequency (tf-idf) weighting is used. The IDF for a term is defined as follows
[23].
Idft,d
log | D |
| {d: t d} |
Eq. 3.1 (e.2)
where
| D | : the total number of documents
: number of documents where the term t appears (i.e.,
As defined in [22], the tf-idf weighting scheme assigns the term t a weight in document d given
by
tf-idft,d = tft,d . idft

In other words, term frequency-inverse term frequency assigns to term t a weight in document d
is
high when term occurs many times within a small number of documents ;
244
lower when the term occurs fewer times in a document, or occurs in many documents;
lowest when the term occurs in virtually all the documents.
Words
abil
abs
enc
Accep
t
1
2
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
5
6
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
10
0.0
0.0
0.0
Access
accoria
account
actio
n
0.0
0.0
0.069434622
98821593
0.120849276
3637866
0.0
0.0
0.071144256
6706304
0.0
0.129932379
10966035
0.079183463
69069837
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
0.0
Row No.
0.0
0.2787299435
038924
0.0
TABLE 3.1 (e): Example set of feature vector with their tf-idf
Now after implementing all the above processing we get each vulnerability description as a
vector, weight that is calculated on using tf-idf formula. Example set of ten rows is shown in table
3.1 (e). This vector form will be used in the scoring and ranking of vulnerabilities.
3.2 Categorization Using Support Vector Machine
Text categorization is the process of categorizing text documents into one or more predefined
categories or classes. Differences in the results of such categorization arise from the feature set
chosen to base the association of a given document with a given category [24]. There are a
number of statistical classification methods that can be applied to text categorization, such as
Nave Bayesian [25], Bayesian Network [25], Decision Tree [26, 27], Neural Network [28], Linear
Regression [29], k-NN [30]. SVM (support vector machine) learning method introduced by [31],
are well-founded in terms of computational science. Support vector machines have met with
significant success in numerous real-world learning tasks [25]. Compared with alternative
machine learning methods including Naive Bayes and neural networks, SVMs achieve
significantly better performance in terms of generalization [32, 33]
SVM classification algorithms, proposed by Vapnik [34] to solve two-class problems, are based
on finding a separation between hyperplanes defined by classes of data, shown in Figure 3.2 (a).
245
FIGURE 6.2.2: Example of SVM hyper plane pattern
This means that the SVM algorithm can operate even in fairly large feature sets as the goal is to
measure the margin of separation of the data rather than matches on features [24]. The SVM is
trained using pre-classified documents.
As explained in [34], for a given set of training data T = {.xi, yi} (i = 1,, m), each data point .xi
Rd with d features and a true label yi Y = {l1, . . . , lk}. In case of binary classifier, label set is Y
= {l1 = 1, l2 = +1}, it classifies data point in positive and negative by finding separating
hyperplane. The separating hyperplane can be expressed as shown in Eq. 3.2 (a).
w .x + b = 0 --------------------------------------------------------------------------Eq 3.2(a)
where w Rd is a weight vector is normal to the hyperplane, operator () computes
the inner-product of vectors w and x and b is the bias. Now we want to choose the wand b to
maximize the margin, or distance between the parallel hyperplanes that are as far apart as
possible while still separating the data. These hyperplanes can be described by the equations
w .x - b = 1--------------------------------------------------------------------------Eq 3.2 (b)

and
w .x - b = -1-------------------------------------------------------------------------Eq 3.2 (c)

In the case that the label set Y = {l1 = 1, . . . , lk = k} and k > 2, a multiclass SVM learning model
is built with two approaches in the proposed framework: multiclass-to binary reduction and
multiclass-optimization methods [10]. In the multiclass-to-binary reduction method, the learning
problem in question is reduced to a set of binary classification tasks and a binary classifier is built
independently for each label lk with the one-against-rest training technique [35]. When more than
246
two classes are involved then regression model is used. regression model builds a classifier
using a regression method which is specified by the inner operator. For each class i a regression
model is trained after setting the label to (+1) if the label equals i and to (-1) if it is not. Then the
regression models are combined into a classification model. Here we are using SVM
classification model, therefore Regression model is combined into SVM. In order to determine the
prediction for an unlabeled example, all models are applied and the class belonging to the
regression model which predicts the greatest value is chosen.
3.3 Identification and Preparation of Training Data
There are number of public and private vulnerability databases that classify vulnerabilities on
different bases like cause, phase of occurrence, product specific etc. the list is shown in Table 3.3
(a). But they all are too generic in nature. The CWE (Common weakness Enumeration) [36] is a
vulnerability database and portal in which each vulnerability is specified with its type, mitigation,
phase of introduction and security property it belongs to. Therefore it is a best place from where
vulnerability data can be collected on the basis of phase of introduction and the security property
it belongs to. Fig 3.3 (a) is screenshot of the CWE window, it is showing a information that each
entry of CWE contains. In this we are interested in only those vulnerabilities that can be mitigated
in the design phase of the software. But in CWE the vulnerabilities are divided into number of
classes and number of examples are given in the description of each class. In order to collect the
training data from CWE, we explore the required security class, then collect vulnerability example
from each class. Almost all the examples that are used in CWE are from CVE (Common
Vulnerability and Exposure).Maximum possible numbers of examples are collected from the CWE
for the training set.
S. No.
1.
2.
3.
4.
5.
6.
7.
8.
9.
Database Name
Common Vulnerability and Exposures
Common Weakness Enumeration
Computer Associates Vulnerability
Encyclopedia
Dragonsoft vulnerability database
ISS X-Force
National Vulnerability Database
Open Source Vulnerability Database
Public Cooperative Vulnerability Database
Security Focus
URL
http://cve.mitre.org/
http://cwe.mitre.org/
http://www3.ca.com/securityadvisor/vulninfo/br
owse.aspx
http://vdb.dragonsoft.com
http://xforce.iss.net/xforce/search.php
http://nvd.nist.gov/
http://www.osvdb.org/
https://cirdb.cerias.purdue.edu/coopvdb/public/
http://www.securityfocus.com/vulnerabilities/
TABLE 3.3 (a): Vulnerability Database with their URLs
247
Vulnerability
is introduced
at the Design
phase of the
SDLC
Examples
from CVE,
that can be
included in
training set of
authentication
vulnerabilitie
s
FIGURE 3.3 (a): CWE Vulnerability Class Description Window
After the exhaustive search of CWE vulnerability classes, the number of vulnerability examples
that are collected, are shown in Table 3.3 (b). While collecting data from CWE, at most care is
taken to include only those vulnerability classes where the time of introduction is specified as
design phase
S. No.
1.
2.
3.
4.
5.
Vulnerability Class
Authentication
Authorization
Audit and Logging
Secure Information Flow
Secure Session Management
Number of
Training data
54
50
36
31
24
TABLE 3.3 (b): Number of training data identified under each class
3.4 Testing and Validation

After the collection of training data, a SVM learning model can be built. But SVM is basically a
binary classifier. As we have multiple classes with |Y |>2, the learning
problem is decomposed into |Y | binary classification tasks with the multiclass-to-binary reduction
method, and subsequently |Y | binary classifiers are built with the one-against-rest training
method that essentially transforms the learning task for category li of Y into a two-class
categorization by treating data points with label li in the training data as positive examples and the
remaining data points as negative examples [10]. We are using Rapidminer tool to implement the
SVM. Regression model is used to classify the data into multiple class. After supplying the data
248
regression model, bootstrap validation is used to validate the classification. Fig 6.2.4 (a) is
showing the screen shot of rapid miner while implementing bootstrap validation.
FIGURE 3.4 (a): Screen shot from Rapidminer Tool, while implementing bootstrap validation
3.5 Bootstrap Validation

The bootstrap family was introduced by Efron and is first described in [37]. In this method, for
given dataset of size n a bootstrap sample is created by sampling n instances uniformly from the
data. There are several bootstrap methods. A commonly used one is the 0.632 bootstrap. As
explained in by Han and Kimber in their book Data Mining: Concepts and Techniques [38], in
this method, suppose we have d tuples. The data set is sampled d/6 times, with replacement,
resulting in the bootstrap samples or the training set of d samples. The data tuple that are not
included in the training, forms the test set. Now the probability for each tuple to be selected is 1/d,
and the probability of not being chosen is (1- 1/d).We have to select d times, so the probability
that a tuple will not be chosen during this whole time is (1-1/d)d. If d is large, the probability
approaches e-1 = 0.368. Thus, 36.8% of tuples will not be selected for training and thereby ends
up in the test set, and the remaining 63.2% will form the training set.
The sampling procedure can be recorded k times, where in each iteration, we can use the current
test set to obtain the accuracy estimate of the model obtained from the current bootstrap sample.
The overall accuracy of the model is then estimated as
k
Acc (M) = (0.632 X Acc (Mi)test-set + 0.368 X Acc(Mi)train_set)
i=1
Eq. 3.5 (a)
where Acc (Mi)test-set is the accuracy of the model obtained with the bootstrap sample i when it is
applied to the test set i. Acc(Mi)train_set is the accuracy of the model obtained with bootstrap
sample i when it is applied to the original set of the data tuples. The bootstrap method works well
with the small data set.
The whole process that is followed in making the classifier is shown in Figure 3.5 (a). the rapid
miner data mining tool is used that have almost all the available data-mining process in the form
of operators. First of all training data is feed to the regression model that is integrated with SVM,
then Apply Model operator is used to apply the created model and performance operator is used
249
to measure the performance of the classifier. As an output, the confusion matrix is be obtained
that will show the accuracy of the classifier.
FIGURE 3.5 (a): Classification Model using Bootstrap Validation
The confusion matrix that is obtained after the application of the classifier is shown in Table 3.5
(a) and the 3D- graphical representation of the confusion matrix is shown in Fig 3.5 (b).
True
True
True
Authorization Others
Pred.
True SecureInformationFlow
True
Audit
True
True SessionClass
and
Authentication management Precision
Logging
Pred.
Authorization
177
96.72%
Pred. Others
139
98.58%
Pred. SecureInformationFlow
131
100.00%
Pred. Audit
and Logging
128
92.09%
Pred.
Authentication
16
189
88.73%
Pred. Sessionmanagement
92
92.93%
95.52%
92.20%
100.00%
Class Recall
98.88% 84.24% 100.00%
Accuracy: 94.52% +/- 1.85% (mikro: 94.48%)
TABLE 3.5 (a): Confusion Matrix
250
FIGURE. 3.5 (b): Confusion matrix in the form of 3D Graph
All most all the classes have class precision value above 90%. The accuracy rate of about 90%
makes the classifier quite accurate (Han and Kamber, 2006). The overall accuracy rate of
developed classifier is 94.5 %.
As shown in Table 3.5 (a), the class precision of authentication class is only 88%, because the
keywords used in the authentication class are common to other classes also. For example the
vulnerability description mainly consist of words like unauthenticated user, not allowed
authenticated user, etc, which actually dont indicate the cause as authentication, but classifier
gets confused due to the frequent use of theses terms in other classes also, which affect the
performance of classifier. But overall accuracy of the classifier is acceptable, which is 94.5%.
4.0 CLASSIFICATION RESULTS

Now using this design level Vulnerability Classification Model the vulnerabilities can be classified
into six classes. In NVD (National Vulnerability Database), total 427 vulnerabilities are identified
as design level vulnerabilities till February 2009. Now in order to classify these vulnerabilities in
our predefined six classes, vulnerabilities first need to be feed in the classifier, then predicted
values can be used for further analysis. After feeding 427 design level vulnerabilities in the model,
the example set of the predicted values that is obtained is shown in Table 4.0 (a). The screenshot
from Rapidminer during the application of the classifier is shown in Fig.4.0 (a) and the final
number of classified vulnerabilities is shown in Fig.4.0 (b). From the classification results it is
clear that out of 427 vulnerabilities that are classified as design level vulnerabilities, 117 are
actually not design level vulnerabilities. From remaining vulnerabilities, Authentication and
authorization related vulnerabilities are most prevailing one, constituting about 53% of total
vulnerabilities.
251
C*
V.N
o.
2
3
4
5
6
Confidence
(Authorizati
on)
0.70382212
6777
0.46653638
8761
0.80392399
6569
0.92738005
8636
0.63117841
7772
0.67407254
0417
1.02248907
1514
Confidence
(Others)
0.74117964
80
1.0
0.56403681
66
0.80057971
24
0.72385838
954
1.03283405
41
0.93112276
32
Confidence
(Secure
Information
Flow)
0.05035211
037
1.04654845
455
0.61342764
113
0.56610700
139
0.99080223
010
1.03255019
721
0.13558666
871
Confidenc
e
(Session
Managem
ent)
Prediction
0.0
Audit and
Logging
0.0
Others
1.0
0.0
Authentica
tion
0.01646968
988
1.0
0.0
Authentica
tion
0.45735700
877
1.0
0.0
Authentica
tion
0.32822280
439
1.0
0.0
Authentica
tion
0.0
Audit and
Logging
Confidence
(Audit and
logging)
1.0
0.54866717
509
0.39066602
401
1.0
Confidence
(Authenticati
on)
0.995248931
575
0.733920948
135
0.563459046
881
TABLE 4.0 (a): Sample dataset from classification result
Percentage of Authentication is 30%, which makes it most important property to be mitigated at

the design phase of the software. Authorization constitute 23.2% of all the vulnerabilities, which
makes it second most important security attribute to be
FIGURE 4.0 (a): Screenshot from rapid miner, while implementing the final model
252
The percentage of audit and logging, secure information flow and session management are
18%, 15% and 12% respectively, which makes them almost equally important.
Vulnerability Class
Count
Percentage
Authentication
Authorization
Audit and Logging
Secure-Information-Flow
Session-management
Others
Total
96
72
56
47
39
117
427
22.48244
16.86183
13.11475
11.00703
9.133489
27.40047
100
Percentage
excluding Others
30.96774
23.22581
18.06452
15.16129
12.58065
0.0
100
TABLE 4.0 (b): Number of vulnerabilities classified under each class
These vulnerability classification data can be used with the severity rating to calculate the risk of
vulnerability occurrence at the design phase.
5.0 CONCLUSION AND FUTURE WORK

As discussed in the previous sections, study of known vulnerabilities is very useful tool for the
developer. Our approach is in the direction of identifying, classifying and learning from known
vulnerabilities. So that these vulnerabilities can be avoided in the next generation of the software.
In available vulnerability databases, there is no information about the vulnerability cause or the
SDLC phase in which they can be removed. Using our proposed classification model, developer
would be able to classify any vulnerability from any vulnerability database. The classification
model will tell the developer whether the vulnerability be mitigated at the design level? If
vulnerability is identified as a design level vulnerability then it will classify the identified
vulnerability in security feature. After knowing the class of security feature, designer can adapt
necessary design patterns that can prevent these vulnerabilities in the new under-developed
software. The accuracy of the classified is found to be satisfactory and it can be used to classify
future vulnerabilities. Classifying exposure leading to access violation class of vulnerabilities is
one of the prompt future works that can be done. The classification results can further be used to
calculate the security risk at the design phase of the software. After risk calculation, mitigation
mechanisms in the form of design patterns can be identified and thus designer will be able to
mitigate security vulnerabilities at the design phase of the software. Another future work that can
be done is the creation of tool that can automate the task of vulnerability classifications. After this
classification our prompt objective will be to identify, analyze and classify design patterns that can
be adapted in order to avoid vulnerabilities in the new software.
REFERENCES
[1]
P.T. Devanbu and S. Stubblebine, Software Engineering for Security: a Roadmap.

International Conference on Software Engineering 2000 special volume on the Future of
Software Engineering, 2000, pp.227-239.
[2]
G. Hoglund and G. McGraw. Exploiting Software: How to Break Code, New York:
Addison-Wesley, 2004
[3]
L. Lowis and R. Accorsi. On a Classification Approach for SOA Vulnerabilities, 33rd

Annual IEEE International Computer Software and Applications Conference. 2009, pp 439444.
[4]
V.C. Berghe, J. Riordan and Piessens A Vulnerability Taxonomy Methodology applied to

Web Services, 10th Nordic Workshop on Secure IT Systems, 2005.
253
[5]
N. Moha. Detection and Correction of Design Defects in Object-Oriented Designs.

Doctoral Symposium, 21st International Conference on Object-Oriented Programming,
Systems, Languages and Application, 2007.
[6]
I.V. Krsul, Software Vulnerability Analysis. Ph.D. Thesis. Purdue University. USA, 1998.
[7]
S. Rehman, and K.Mustafa. Software Design Level Security Vulnerabilities, International

Journal of Software Engineering, 4 (2). 2011.
[8]
T. Joachims. Text categorization with support vector machines: learning with many
relevant features. 10th European Conference on Machine Learning. 1998.
[9]
J. A. Wang, and M. Guo. OVM: An Ontology for Vulnerability Management. 7th Annual
Cyber Security and Information Intelligence Research Workshop.Tennessee, USA. 2009.
[10]
Z. Chen, Y. Zhang, and Z. Chen A Categorization Framework for Common Computer

Vulnerabilities and Exposures. Computer Journal Advance Access, 2009. Available:
http://comjnl.oxfordjournals.org/ cgi/content/abstract/bxp040.
[11]
P.H. Meland, and J. Jensen. Secure Software Design in Practice. Third International
Conference on Availability, Reliability and Security. 2008.
[12]
Y. Li, H.S. Venter, and J.H.P Eloff. Categorizing vulnerabilities using data clustering
techniques, Information and Computer Security Architectures (ICSA) Research Group.
2009.
[13]
N.H.Pham, T.T Nguyen, H.A Nguyen,., X.Wang, , A.T. Nguyen, and T.N Nguyen.
Detecting Recurring and Similar Software Vulnerabilities, International Conference of
Software Engineering. Cape Town, South Africa. 2010.
[14]
D. Byers, S. Ardi, , N. Shahmehri and C. Duma. Modelling Software Vulnerabilities with

Vulnerability Cause Graphs. 22nd IEEE International Conference on Software
Maintenance. , 2006.
[15]
V. Sridharan, and D.R. Kaeli . Quantifying Software Vulnerability. Workshop on Radiation

effects and fault tolerance in nanometer technologies, Ischia, Italy, 2008.
[16]
Y.Wu, R.A. Gandhi, and H. Siy. Using Semantic Templates to Study Vulnerabilities
th
Recorded in Large Software Repositories. 6 International workshop on software
Engineering for secure system, Cape Town, South Africa. 2010.
[17]
G. Grefenstette and P. Tapanainen. What is a Word, What is a Sentence? Problems of

Tokenization. 3rd Conference on Computational Lexicography and Text Research . 1994,
pp. 79-87.
[18]
C. Fox. Lexical Analysis and Stoplist-Data Structures and Algorithms. New York: PrenticeHall. 1992.
[19]
M. F. Porter. Snowball: A string processing language for creating stemming algorithms in

information retrieval, 2008. Available: http://snowball.tartarus.org.
[20]
Lemur Project (2008). The Lemur Toolkit: For Language Modeling and Information
Retrieval, 2008. Available: http://www.lemurproject.org.
[21]
M. Braschler and B. Ripplinger, How Effective is Stemming and Decompounding for

German Text Retrieval. Information Retrieval, 7, 2003, pp.291316.
254
[22]
C.D. Manning, P. Raghavan, and H. Schtze. Introduction to Information Retrieval,

Cambridge University Press. 2008.
[23]
A. Rajaraman, and J.D. Ullman, Mining of Massive Datasets. 2010. Available:

http://infolab.stanford.edu/~ullman/mmds/ch1.pdf
[24]
A. Basu, C. Walters, M. Shepherd. Support vector machines for text categorization. 36th
Annual Hawaii International Conference,2003
[25]
T. Joachims. A probabilistic analysis of the Rocchio algorithm with TFIDF for text
categorization, 14th International Conference on Machine Learning. 1997.
[26]
J.R. Quinlan. Programs for machine learning. San Francisco: Morgan Kaufmann
Publishers.1993.
[27]
S. M. Weiss, C. Apte, F.J. Damerau, D.E. Johnson, F.J. Oles, T., Goetz, T. Hampp.
Maximizing text-mining performance. IEEE Intelligent Systems Magazine, 1999.
[28]
E. Wiener, J. O. Pederson, A.S. Weigend. A neural network approach to topic spotting,

4th Annual Symposium on Document Analysis and Information Retrieval. 1995.
[29]
Y. Yang and , J.O. Pederson. A comparative study on feature selection in text

categorization. International Conference on Machine Learning. 1997.
[30]
Y. Yang. An evaluation of statistical approaches to text categorization. Journal of

Information Retrieval. 1 (2). 1999.
[31]
V. Vapnik,. The Nature of Statistical Learning Theory. Berlin: Springer. 1995.
[32]
C. Burges. "A tutorial on support vector machines for pattern recognition. Data Mining and
Knowledge Discovery, 2, 1998, pp. 1-47.
[33]
J.T.K. Kwok. Automated Text Categorization Using Support Vector Machine. International
Conference on Neural Information Processing, 1998.
[34]
V. Vapnik. Statistical Learning Theory. New York: John Wiley and Sons. 1998.
[35]
T. Hastie, and R. Tibshirani, Classification by pair wise coupling. Ann. Statist, 26, 1998,
pp. 451471.
[36]
CWE (Common Weakness Enumeration). Available: http://cwe.mitre.org/
[37]
B. Efron. Estimating the error rate of a prediction rule: Improvement on cross-validation.

Journal of the American Statistical Association, 78, 1983. pp.316-331.
[38]
J. Han, and M. Kamber Data Mining: Concepts and Techniques. San Francisco: Morgan
Kaufmann Publisher, 2006.
255

Software Design Level Vulnerability Classification Model

Uploaded by

Copyright:

Available Formats

Software Design Level Vulnerability Classification Model

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Software Design Level Vulnerability Classification Model

Uploaded by

Copyright:

Available Formats

Shabana Rehman & Khurram Mustafa