IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 14, No. 1, February 2025, pp. 337~349

ISSN: 2252-8938, DOI: 10.11591/ijai.v14.i1.pp337-349  337

SQL-CB-GuArd: a deep learning mechanism for structured

query language injection attack detection

AsifIqbal Sirmulla, Prabhakar Manickam

School of Computer Science and Engineering, Reva University, Bangalore, India

Article Info ABSTRACT

Article history: Structured query language (SQL) injection attacks, which take advantage of
input field vulnerabilities to introduce malicious code into database queries,
Received Mar 4, 2024 are a serious danger to database-driven programs and systems. Intruders can
Revised Jul 3, 2024 now alter, recover, or remove sensitive data because of illegal access. Strong
Accepted Jul 26, 2024 artificial intelligence (AI) based security solutions are required to reduce
SQL injection threats, as these assaults' significance highlights. This study's
main goal is to create automated AI-based techniques that can identify
Keywords: structured query language injection attack (SQLIA) in real time eliminating
the need for human intervention. Although machine learning (ML) and deep
Attention mechanism learning-based techniques have received a lot of interest in this field, ML-
Bidirectional long short-term based techniques have problems with accuracy and false negatives. Deep
memory learning (DL) is therefore commonly used in these text data processing and
Convolutional neural network natural language processing (NLP) applications. We have introduced a
Deep learning hybrid DL approach for SQLIA detection in this paper. The pre-processing
Gated recurrent unit step performs decoding, generalization, and tokenization to improve the
Natural language processing learning performance. The proposed approach uses combination of
Structured query language convolutional neural network (CNN), bidirectional long short-term memory
injection attack (Bi-LSTM), gated recurrent unit (GRU) with attention mechanism. The
combination helps to improve the pattern learning capacity. The proposed
approach is validated on publically available data and experimental analysis
reported that the proposed SQL-CB-GuArd achieves better accuracy of
SQLIA detection.
This is an open access article under the CC BY-SA license.

Corresponding Author:
AsifIqbal Sirmulla
School of Computer Science and Engineering, REVA University
Bangalore, India
Email: [email protected]

1. INTRODUCTION
In today's world, information technology has become essential in various aspects of our lives,
including communication, transportation, and banking. While these advancements bring many benefits, they
also open doors for malicious actions in cyberspace [1]. Similarly, the modern web-based technologies and
cloud based systems got very popular due to their diverse range of applications [2], [3]. Therefore, these
systems are easily accessible for every individual. However, these systems raise serious concerns over
security of the systems. Several studies have reported the increase in cyberattacks. For example, in 2020,
there were about 4,000 cyberattacks recorded daily worldwide. Cybersecurity ventures estimated that by
2021, cybercrime would cost the world economy $6 trillion annually, compared to $3 trillion in 2015 [4].
Ransomware attacks have seen a significant increase, with a 150% rise in 2020 compared to 2019 [5].
Similarly, phishing attacks remain a major threat, with over 240,000 unique phishing websites detected each
month in 2020 [6]. The anti-phishing working group reported a 22% increase in phishing attacks in the first

Journal homepage: http://ijai.iaescore.com

338  ISSN: 2252-8938

half of 2021 compared to the same period in 2020. These attacks often lead to data breaches, which have
become a significant concern. In 2020, over 37 billion records were exposed in breaches, marking a 141%
increase from 2019 [7], [8]. Based on IBM's cost of a data breach report [9], the average time to detect and
remediate a data breach is 280 days. Hence, it's crucial to protect cyber infrastructures from harmful threats.
Similarly, the IT infrastructure also has noticed a tremendous growth and its dependency on the
technology has increased drastically. Therefore, all most all applications which are used in daily life are
utilizing web-based system to operate them efficiently. As discussed before these web-based applications are
easily accessible to individuals therefore it brings various security challenges because of illegal access [10].
When users use web apps or portals, user-generated data gets stored within these platforms or their respective
backend databases. These databases are overseen by database management systems (DBMS), which execute
request queries scripted in the form of structured query language (SQL). Both the frontend app and backend
db are susceptible to various forms of assaults due to their internet accessibility. Amongst the various
vulnerabilities outlined by OWASP, structured query language injection attack (SQLIA) rank among the top
10 [11]. Recent trends indicate a staggering increase in attack rates, exceeding 300% over the past decade,
with attackers employing sophisticated tactics like obscured or cipher codes to bypass security systems [12].
SQLIA occurs when an SQL query is injected or inserted via a client-side data input field into the online app
or portal [13]. The injected query can be malicious, potentially resulting in database modification, identity
spoofing, sensitive information leakage, and repudiation issues if successfully executed [14]. Therefore,
several methods have been introduced to address the SQLIA issue such as input validation, parameterized
queries, web application firewalls (WAF), input filtering and sanitization, database activity monitoring
(DAM), error handling and logging, and black box and white box testing [15]. Table 1 shows a comparative
analysis where the advantages and drawbacks of these models are briefly described.

Table 1. Traditional methods of attack detection

Method Description Drawback
Input Putting in place stringent input validation procedures to guarantee Fails in complex applications with
validation that data submitted by users satisfies expectations regarding numerous input fields.
format, length, and data type.
Parameterized Using prepared statements or parameterized searches when It requires significant changes in existing
queries interacting with databases. By separating SQL code from user codebase and implementation can be error
input, these techniques stop injected SQL instructions from being prone
executed.
WAF It keeps an eye on and filters HTTP traffic going between the It may lead to generate the false positive
internet and an online portal. Based on specified rulesets, WAFs and false negative leading to block the
are able to identify and prevent SQL injection activities. legitimate traffic or allowing the
malicious traffic
Input filtering It removes or neutralize potentially harmful characters or Sometimes it fails to catch all variations
and sanitization commands from user input before processing it. of input malicious traffic
DAM It offers tools for real-time DAM and analysis. It can detect It leads to produce high volume of alerts
anomalous SQL queries or behaviors indicative of a potential
SQLIA.
Error handling It implements error handling within the application to capture and Large log volumes becomes time-
and logging log any SQL-related errors. Analyzing error logs can help identify consuming
potential SQL injection attempts.

Presently, machine learning (ML) based systems have gathered huge attention in various
applications such as image processing, video processing biomedical domain, textual data analysis and natural
language processing (NLP) tasks. ML based methods have been considered as promising method to detect
the SQL attacks and other intrusion successfully. Nofal and Amer [16] presented a combined approach where
neural network and fuzzy logic methods were used for SQLIA detection. Sheykhkanloo [17] presented in
neural network based model which uses uniform resource locator (URL) generator, URL classifier, and NN
model. Similarly, support vector machine (SVM), decision tree (DT), and k-nearest neighbors (KNN)
classifiers are also employed for SQLIA detection. However, the reliability and generalization of these
methods remains a challenging task. Therefore, deep learning (DL) based methods have been adopted to
improve the performance. the literature review presents the brief discussion about these methods.
In this work, we present a hybrid DL based solution for SQLIA detection. The proposed work uses
combination of data pre-processing, vectorization, and implementation of DL model. The data pre-processing
stage implements decoding, generalization, and tokenization. In next stage vectorization is performed.
Finally, the DL model is implement which uses convolutional neural network (CNN), bidirectional long
short-term memory (Bi-LSTM), and gated recurrent unit (GRU) with attention mechanism.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 337-349

Int J Artif Intell ISSN: 2252-8938  339

Rest of the article is arranged in following section. Section 2 presents brief literature review about
standard methods. Section 3 showcases the proposed model. Section 4 presents the outcome of proposed
model along with the comparative analysis with existing methods. Finally section 5 presents the concluding
remarks about the work.

2. LITERATURE SURVEY
This section provides a quick overview of the literature on current techniques for detecting SQLIA
through the use of DL and ML techniques. Vartouni et al. [18] focused on enhancing WAF using deep ML
algorithms, specifically focusing on anomaly detection. Initially, attributes are constructed from HTTP data
using two models: n-gram and one-hot. Subsequently, employing auto-encoder long short-term memory
networks (AE-LSTM) as an unsupervised DL technique, informative features are extracted and reduced.
Finally, an ensemble isolation forest is utilized to train the classifier exclusively on normal data.
Thalji et al. [19] presented DL based model by using recurrent neural network (RNN) based pattern learning
approach. Moreover, it implements autoencoder based system along with RNN. The encoder module
considers the input data and compresses the data to lower dimension. This encoded data then fed into the
decoder module which reconstructs the data. Finally, long short-term memory networks (LSTM) and dense
layer based RNN model is employed for classification.
Gandhi et al. [20] reported that traditional prevention methods, such as rule-based matching, have
limitations in detecting a wide range of SQLIA. One major challenge is the constant evolution of malicious
SQL queries by hackers. To address this, leveraging ML algorithms for predicting SQLIA proves promising.
This paper proposes a hybrid CNN-Bi-LSTM method for identification of SQLIA. the complete architecture
utilizes several layers such as embedding layer, convolution layer, and max pooling with Bi-LSTM. Finally,
dense and output layers are used to obtain the final outcome. Sun et al. [21] reported accuracy and false
negative issues of existing methods. Recent ML research on SQLIA identification has focused mainly on
extracting the features. However, the efficacy of identification relies heavily on the accurateness of features
extracted, which may not adequately address more complex SQLIA. To tackle these challenges, a novel
approach to SQLIA detection, combining an enhanced TextCNN and LSTM is presented in this work. In
order to acquire local features, this technique first vectorizes samples inside the corpus and then uses an
enhanced TextCNN. The sequence information present in the samples is then captured using a Bi-LSTM
network. Understanding that LSTM is less successful with longer sequences, it employs an attention
mechanism to shorten the gap between any two words in the sequence to one. The feature selection procedure
also incorporates pre-trained word vector features acquired using bidirectional encoder representations from
transformers (BERT) for transfer learning. A mixed ML model based on LSTM and SVM was reported by
Fang et al. [22]. Initially the likelihood ratio is used, and then SQL tokenization. This produces a word vector
model in SQL. This model is given into the LSTM model as a train set, and it classifies the patterns as
authentic and injected queries. Kherbache et al. [23] talked about how feature selection, which involves
choosing the most critical qualities to increase efficiency and reduce false positives, is seen to be a crucial
step in anomaly-based intrusion detection. Authors offered a mixed agglomerative hierarchical clustering
approach with SVM classification based on this idea. Based on their commonalities, features are categorized
and feature selection is carried out according to variance. A CNN-based model for online security was
provided by Jemal et al. [24], who also examined the significance of the hyper-parameters included in CNN
models.

3. PROPOSED MODEL
This section presents the proposed DL based model for SQLIA detection. As discussed before, the
traditional methods have several disadvantages which are further addressed by DL based methods. However,
the ever evolving process of attackers affect the SQLIA detection performance. In order to overcome the
issues of existing models, we introduced a new hybrid DL model which utilizes four different concepts to
improve the performance. The overview of these concepts is given as follows:
− CNN: it is a class of DL mechanism commonly employed for image and non-image (including textual)
data. The CNN based method utilizes learnable filters (also called kernels or convolutional kernels) to
convolve over the input data, fetching local patterns and features.
− LSTM: It is a kind of RNN architecture intended to recognize long-term relationships in sequential data
and solve the vanishing gradient issue. LSTM networks are appropriate for jobs where comprehending
context over long distances is critical because they are able to selectively store or discard info over
lengthy durations. They are now a common option for many sequence modeling tasks, particularly those
involving time-series data or jobs involving NLP, such text creation, sentiment analysis, and machine
translation.
SQL-CB-GuArd: a deep learning mechanism for Structured query language … (AsifIqbal Sirmulla)
340  ISSN: 2252-8938

− GRU: Akin to the LSTM, the GRU is another kind of RNN design that is intended to handle the
vanishing gradient issue and identify long-term relationships in sequential data. GRUs are more effective
in training and deploying than LSTMs since they have fewer parameters and need less computing power.
− Attention mechanism: When making predictions, neural networks, especially sequence-to-sequence
models, employ this technique to concentrate on pertinent portions of the input data. It enables the model
to focus on specific segments of the input sequence, giving each segment varying degrees of significance
according to how relevant it is to the present forecast.
The proposed architecture is designed based on the combination of these architectures. Figure 1
depicts the complete proposed architecture for SQLIA detection. It includes four different phases where first
phase includes data pre-processing which includes decoding, generalization, and tokenization, next phase
performs vectorization, phase III uses this vectorised data through the proposed classifier model and finally,
the classification outcomes are analayzed.

Figure 1. Proposed SQL-CB-GuArd (SQL-CNN Bi-LSTM Gated Recurrent Attention Mechanism)

3.1. Data pre-processing

This step performs several tasks as the data pre-processing phase such as data decoding,
generalization, and tokenization. Data decoding includes generalization and tokenization tasks. In next phase,
vectorization converts the dataset into numerical vectors.

3.1.1. Decoding
The first stage of this work performs data pre-processing where data decoding, generalization and
tokenization tasks are performed. The data decoding refers to the process of converting encoded or sanitized
input data back to its original form before further analysis or processing. The decoding process involves
reversing the encoding or sanitization applied to the input data, so that the original input can be analysed for
signs of SQLIA. For example, if user input has been URL-encoded or hyper text markup language (HTML)-
encoded to prevent SQL injection, the defence mechanism would need to decode the input data to recover the
original user input before passing it to the SQL injection detection model. Similarly, if special characters
have been escaped or replaced with safe alternatives, you would need to reverse this process to obtain the
original input data. Table 2 shows the example of URL encoded and HTML encoded inputs by user and their
corresponding decoded output. This step converts the complex hyperlinks in simplified forms.

Table 2. Example of decoding URL encoding and HTML decoding

Input Decoded Operations
URL encoding “Hello%20World%21” Hello World! %20 replaced with space character
HTML decoding "&lt;script&gt;alert('XSS <script>alert('XSS the HTML entities &lt; and &gt; are
Attack!')&lt;/script&gt;" Attack!')</script> replaced with < and >, respectively,

Int J Artif Intell, Vol. 14, No. 1, February 2025: 337-349

Int J Artif Intell ISSN: 2252-8938  341

3.1.2. Generalization
In this stage, the data generalization operation is performed where the main aim is to minimize the
redundant and irrelevant distribution. This generalization process includes several steps such as:
− Handling the URL: generally, the input data consist of several URLs which are replaced with a
placeholder string such as “https://website”. This step ensures the mitigating the impact of specific URLs,
especially if the URLs are not relevant to the task at hand or if they contain sensitive information.
− Replacing numbers: Numbers in the data are replaced with a placeholder value, such as '0'. This step is
often done to generalize numerical data and focus on the structural aspects of the data rather than the
specific numeric values. It can help reduce the dimensionality of the data and mitigate the influence of
numeric outliers.
− Removing extra unique qualities: Certain unique qualities in the data, such as control characters or blank
characters, are removed. Control characters, such as newline characters or tab characters, can introduce
noise or artifacts in the data that may interfere with analysis or modeling. Removing these characters
helps in standardizing the data and making it more responsive to processing. For example, the SQL
injection query is "SELECT * FROM users WHERE username = 'admin' --' AND password = 'password'"
and after removing unique qualities it becomes "SELECT * FROM users WHERE username = 'admin'
AND password = 'password'".

3.1.3. Tokenization
Tokenization plays important role in this domain of SQLIA detection. This process breaks down the
SQL queries or input into various smaller and meaningful units which are called as tokens. Each token
represents the word, symbol or other elements in the query. Moreover, it helps to extract the significant
features for analysis. The tokenization performs several steps such as splitting queries into token, removing
comments, extracting features from tokens such as token frequencies, n-grams (sequences of tokens),
syntactic patterns, or semantic information derived from the tokens, and building the vocabulary. However,
this work uses different scripting languages therefore the feature extraction is performed based on input data.
Thus, the proposed tokenization approach performs several steps to accomplish the tokenization process such
as identifying starting and ending labels, identification of windows events and function names, assigning
unique tokens, analysing the vocabulary list and predefined delimiter replacement. Table 3 shows the
complete overview of tokenization process.
− Identification of starting and ending labels: this process involves identifyting the starting and ending
labels of various elements in web content such as identifying the HTML tags, event handles, JavaScript
function etc. For example, in HTML, tags are enclosed within angle brackets (< >), and JavaScript
functions are defined using the function keyword followed by a pair of parentheses and curly braces ({ }).
− Identification of window event and function names: in this phase, the proposed tokenization process
involves identifying the certain features related to the windows events and function names which are
commonly used in XSS attacks. These event handlers are 𝑜𝑛𝑐𝑙𝑖𝑐𝑘, 𝑜𝑛𝑚𝑜𝑢𝑠𝑒𝑜𝑣𝑒𝑟, and 𝑜𝑛𝑙𝑜𝑎𝑑.
Similarly, the function names are “eval”, “alert”, “prompt”.
− Unique token assignment: once the starting and ending labels, as well as Windows event and function
names, are identified, unique tokens are assigned to each of these features. These tokens serve as
representations of the features and are used to create a vocabulary list.
− Checking tokens in vocabulary list: in this step the tokens obtained from the input data is checked against
the vocabulary list which consist of predefined tokens to represent the identified attributes. If the obtained
token matches with any of the predefined attribute, then it is retained otherwise it is replaced with the
predefined delimiter or placeholder token.

3.2. Vectorization
This is the next step after pre-processing which is used to represent the SQL queries or input data as
numerical vectors. This representation is helpful for ML and DL methods to perform classification tasks.
Several methods have been introduced to perform the vectorization such as bag of word, term frequency-
inverse document frequency (TF-IDF), and word embeddings etc. In this work, we employed word2vec
model for word embedding. After completing the pre-processing tasks, this word embedding model is
applied. According to this, initially vocabulary is created which is constructed by selecting the most common
words from tokenized data. This vocabulary consists of unique words present in the corpus and serves as the
basis for generating word embeddings. Typically, words are represented by numerical indices in the
vocabulary.
After vocabulary, the 𝑤𝑜𝑟𝑑2𝑣𝑒𝑐 model need to be trained. Word2Vec utilizes a neural network
architecture, specifically either the “continuous bag of words (CBOW)” model or the “skip-gram model”, to
learn distributed representations of words based on their contexts. This stage involves training the neural

SQL-CB-GuArd: a deep learning mechanism for Structured query language … (AsifIqbal Sirmulla)
342  ISSN: 2252-8938

network to estimate a word's probability provided its surrounding words (for CBOW), or given a target word,
its adjacent words (for skip-gram). This training process involves adjusting the neural network's parameters
(word vectors) to minimize prediction errors.
Let 𝑊 represent the Word2Vec model, which consists of a neural network with parameters 𝜃. The
training process aims to minimize a loss function ℒ by adjusting the parameters 𝜃. Mathematically, this can
be represented as (1):
1
𝑚𝑖𝑛𝜃 . ∑𝑁
𝑖=1 ℒ (𝑊(𝑇𝑖 ), 𝑡𝑖𝑐 ) (1)
𝑁

Where 𝑊(𝑇𝑖 ) is the output of the Word2Vec model for the input token sequence 𝑇𝑖 , and 𝑡𝑖𝑐 is the context
token for the central word in the sequence.
Following training, every word in the vocabulary has an embedding, or dense vector representation,
in a continuous vector space. This is known as the Word2Vec model. Similar words are represented by
adjacent vectors in the vector space as a result of these embeddings, which reflect the semantic links between
words. The vector embeddings generated by Word2Vec are mapped with the vocabulary created earlier. Each
word in the vocabulary is associated with its corresponding vector representation in the continuous vector
space. For example, let us consider that we have a SQL query given as:

“SELECT * FROM users WHERE username='admin'

This query is processed through the Word2Vec model to produce embeddings for each word in the query.
The outcome of this Word2Vec model is presented as:

Word2Vec(‘SELECT’)=[0.1, 0.2, 0.3],

Word2Vec(‘*’)=[0.2, 0.3, 0.4]
Word2Vec(‘FROM’)=[P0.3, 0.4, 0.5]
Word2Vec(‘users’)=[0.4, 0.5, 0.6]
Word2Vec(‘WHERE’)=(‘0.5, 0.6, 0.7’)
Word2Vec(‘username’)=(‘0.6, 0.7, 0.8’)
Word2Vec(‘=’)=(‘0.7, 0.8, 0.9’)
Word2Vec(‘admin’)=(‘0.8, 0.9, 1.0’)

Further, in order to obtained the final vector representation of given SQL query, we perform
averaging on the Word2Vec embeddings with the help of tokens, which is expressed as:

SQL_Query_Vector = (Word2Vec('SELECT') + Word2Vec ('*') + Word2Vec('FROM') + Word2Vec('users')

+ Word2Vec('WHERE') + Word2Vec('username') + Word2Vec ('=') + Word2Vec('admin'))/8
=([0.1, 0.2, 0.3]+[0.2, 0.3, 0.4]+[0.3, 0.4, 0.5]+[0.4, 0.5, 0.6]+[0.5, 0.6, 0.7]+[0.6, 0.7, 0.8]+[0.7, 0.8, 0.9]+
[0.8, 0.9, 1.0])/8
=[0.475, 0.575, 0.675] is the final SQL_Query_Vector. This vector captures the semantic meaning
of the query in a continuous vector space, which can be used as input to proposed hybrid DL model. The
proposed architecture is discussed in next section.

Table 3. Tokenization process

Tokenization Input (potentially with SQLIA) Tokenized Output Description
Starting and <input type="text" name="username" ['<', 'input', 'type', '=', '"text"', <input> is identified as a starting
end label value="' OR 1=1; --"> 'name', '=', '"username"', 'value', label, and its corresponding ending
identification '=', "' OR 1=1; --'", '>'] label > is also recognized.
window <script> ['<', 'script', '>', 'var', 'username', the JavaScript function
event and var username = 'admin'; '=', "'admin'", ';', 'var', 'password', executeQuery and the SQL query
function var password = 'password'; '=', "'password'", ';', 'function', string within the authenticate
names function authenticate() { 'authenticate', '(', ')', '{', 'var', function are identified as features
identification var query = "SELECT * FROM users 'query', '=', '"SELECT * FROM of interest. The SQL query string
WHERE username='" + username + "' users WHERE username=\'" + concatenates user inputs username
AND password='" + password + "';"; username + "\' AND password=\'" and password, making it
executeQuery(query); + password + "\';"', ';', vulnerable to SQL injection.
} 'executeQuery', '(', 'query', ')', ';',
</script> '}', '<', '/script', '>']
Unique <input type="text" name="search" ['<', 'input', 'type', '=', '"text"', he SQL injection payload ); DROP
token value="); DROP TABLE users; --"> 'name', '=', '"search"', 'value', '=', TABLE users; -- is treated as user
assignment "); DROP TABLE users; --", '>'] input and tokenized accordingly.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 337-349

Int J Artif Intell ISSN: 2252-8938  343

3.3. Proposed SQL-CB-GuArd (SQL-CNN Bi-LSTM gated recurrent attention mechanism) for
SQLIA detection
This section describes the proposed DL based solution for SQLIA detection. The proposed
architecture is developed by using the combination of CNN, Bi-LSTM and GRU models to obtain the final
classification model. The complete detail of this model is described as follows.

3.3.1. Convolutional neural network

The CNN model plays important role in extracting the features from the input data. For obtaining
the features from the input data, it uses several kernels to conduct convolution operations. To increase the
network's flexibility, the kernel is made up of bias term and trainable weight coefficients. This is the
fundamental layer of the network and the output of this layer is processed further where ReLU activation
function is performed which is expressed as (2):

𝑓(𝑥) = max(0, 𝑥) (2)

The ReLU activation helps to mitigate the vanishing gradient and improves the training process. In
next step, max pooling operation is performed. The pooling operation downsamples the duplicate data and
focuses on identifying the invariance in the network. here, two pooling operations are performed such as max
pooling and average pooling. The max pool helps to select the maximum value and average pool helps to
select the average value as output. The pooling operations helps to reduce the spatial dimension of feature
maps resulting in the increase of network efficiency. Finally, the fully connected layer is employed to
establish the relation between input and output. The outcome of CNN model can be expressed as (3):

𝑦𝑖 = 𝐶𝑁𝑁(𝑥𝑖 ) (3)

According to the SQLIA model, we consider that the input sequence is denoted by 𝑋 with
dimension (𝑇, 𝐷) where 𝑇 is the sequence length and 𝐷 is the embedding dimension. Let 𝑍 𝑐 represents the
output feature map with the dimension (𝑇 ′ , 𝐹) after applying the CNN and max pooling where 𝑇 ′ is the
sequence length obtained by applying pooling and 𝐹 represents the number of filter. Thus, the convolution
operation can be represented as (4):

𝑍 (𝑐) = 𝑅𝑒𝐿𝑈(𝐶𝑜𝑛𝑣(𝑋)) (4)

3.3.2. Bidirectional long short-term memory

The outcome of CNN model is fed into the Bi-LSTM model. This model is based on the working of
RNN model. However, the traditional RNN models suffer from the issue of vanishing gradient problem. The
basic LSTM model consist of three components called as input, forge and output gates as depicted in Figure 2.
The main aim of LSTM model is to regulate the states of these gates to improve the learning performance. The
forget gate 𝑓𝑡 is used to ensure whether to keep the information of previous state (𝑐𝑡 − 1) or not. The input
gate (𝑖𝑡 ). Similarly, the input gate (𝑖𝑡 ) determines the extent to which information from the input text (𝑥𝑡 ) and
the previous hidden state (ℎ𝑡 − 1) should influence the updating of the cell state. Its output can be either 0 or 1.
The value of ct represents the newly generated cell state resulting from computational operations involving
𝑐𝑡 − 1, 𝑓𝑡 , and 𝑖𝑡 . Meanwhile, the output gate (𝑜𝑡 ) regulates the transmission of information from the current
cell state to the hidden state, with its value also being either 0 or 1. These operations can be expressed as (5):

𝑓𝑡 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 (𝑊𝑓𝑥 𝑥𝑡 + 𝑊𝑓ℎ ℎ𝑡−1 + 𝑏𝑓 )

𝑖𝑡 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝑊𝑖𝑥 𝑥𝑡 + 𝑊𝑖ℎ ℎ𝑡−1 + 𝑏𝑖 )
𝑐𝑡 = 𝑐𝑡−1 ⨀𝑓𝑡 + 𝑖𝑡 tanh(𝑊𝑐𝑥 𝑥𝑡 + 𝑊𝑐ℎ ℎ𝑡−1 + 𝑏𝑐 )
𝑜𝑡 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝑊𝑜𝑥 𝑥𝑡 + 𝑊𝑜ℎ ℎ𝑡−1 + 𝑏0 )
ℎ𝑡 = 𝑜𝑡 ⨀ tanh(𝑐𝑡 ) (5)

Where 𝑥𝑡 ∈ 𝑅𝑛 represents the input vector, 𝑊 ∈ 𝑅𝑣∗𝑛 , 𝑏 ∈ 𝑅𝑣 and 𝑛, 𝑣 represents the input vector dimension
and number of words in the vocabulary.
In the LSTM network, information propagates solely in a forward direction, implying that the state
at time 𝑡 is influenced only by information preceding 𝑡. However, to fully capture the semantic context of an
input review, subsequent information is equally significant as previous ones. Thus, for a more comprehensive
representation of contextual information, the Bi-LSTM model is utilized. The Bi-LSTM model consists of
two LSTM networks and has the capability to process input reviews in both forward and backward
directions. Therefore, we adopt the Bi-LSTM model in this work. The Figure 3 depicts the Bi-LSTM layer.
SQL-CB-GuArd: a deep learning mechanism for Structured query language … (AsifIqbal Sirmulla)
344  ISSN: 2252-8938

Figure 2. LSTM model

Figure 3. Bi-LSTM architecture

The forward and backward LSTM processes are expressed as (6):

⃗ 𝑡 = 𝐿𝑆𝑇𝑀(𝑥𝑡 , ℎ
ℎ ⃗ 𝑡−1 )
⃖⃗𝑡 = 𝐿𝑆𝑇𝑀(𝑥𝑡 , ℎ
ℎ ⃖⃗𝑡+1 ) (6)

Finally, the output of Bi-LSTM is summarized by concatenating the forward and backward states as (7):

⃗ 𝑡, ℎ
ℎ𝑡 = [ℎ ⃖⃗𝑡 ] (7)

According to the proposed model, the 𝐻 (𝑙) represents the output hidden states obtained from Bi-LSTM
model. The dimension of 𝐻 (𝑙) are given as (𝑇 ′ , 2𝐻) where 𝐻 represents the number of hidden units. The
Bi-LSTM operation can be expressed as (8):

𝐻 (𝑙) = 𝐵𝑖𝐿𝑆𝑇𝑀(𝑋) (8)

3.3.3. Gated recurrent unit and attention model

Along with CNN and Bi-LSTM, we incorporate the GRU model which plays important role in
sequence modelling tasks. In a typical GRU cell, there are two gates: the reset gate (r) and the update gate
(z). Similar to an LSTM cell, the computation of the hidden state output at time t involves the hidden state at
time t-1 and the input time series value at time t can be expressed as (9):

ℎ𝑡 = 𝑓(ℎ𝑡−1 , 𝑥𝑡 ) (9)

According to the proposed model, the output of GRU can be expressed as (10):

Int J Artif Intell, Vol. 14, No. 1, February 2025: 337-349

Int J Artif Intell ISSN: 2252-8938  345

𝐻 𝑔 = 𝐺𝑅𝑈(𝐻 (𝑙) ) (10)

In order to improve the overall performance of SQLIA, we also incorporate the attention mechanism
where 𝛼 represents the attention weights which are computed with the help of hidden states 𝐻 (𝑔) . The
attention weights can be computed as (11):

𝛼 = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥 (𝐿𝑖𝑛𝑒𝑎𝑟(𝐻 (𝑔) )) (11)

With the help of hidden states, the context vector can be obtained by estimating the weighted sum of these
states, which can be expressed as (12):
′ (𝑔)
𝐶 = ∑𝑇𝑡=1 𝛼𝑡 . 𝐻𝑡 (12)

The overall architecture of proposed model is represented in Figure 4 which shows the layer, its parameters
and connection to another layer.

Figure 4. overall architecture of proposed model

SQL-CB-GuArd: a deep learning mechanism for Structured query language … (AsifIqbal Sirmulla)
346  ISSN: 2252-8938

4. RESULTS AND DISCUSSION

This section presents the results of the suggested technique and makes a comparison between the
acquired performance and several current SQLIA detection schemes. The first subsection presents the
overview of dataset used including performance measurement parameters. Next subsection describes the
obtained performance, and finally a comparative analysis is presented.

4.1. Dataset details and performance measurement parameters

The performance of proposed model is validated on publically available datasets which are obtained
from GitHub and Kaggle websites. Table 4 presents the dataset details. In this table, five different CSV
dataset files are used which are named as dataset 1, dataset 2, dataset 3, dataset 4, and dataset 5 which are
finally combined together to generate the final dataset.

Table 4. Dataset details

Dataset SQLi Query Legitimate Query
Dataset 1 158 994
Dataset 2 1594 994
Dataset 3 5433 994
Dataset 4 22559 19009
Dataset 5 29379 20002

Similarly, we utilized a publicly available SQL query dataset [21] for our research experiments. The
dataset comprises three files: ‘‘sqli.csv,’’ ‘‘sqliv2.csv,’’ and ‘‘SQLiV3.csv,’’ which we combined. For initial
preprocessing, we encoded the target labels ‘Attack’ (1) and ‘Benign’ (0). A PC equipped with an Intel i7
10th GEN CPU (2.20 GHz), 136 GB of RAM, and an 8 GB NVIDIA graphics card was used for all testing.
Tensorflow 1.4.1 is used as a backend calculation, Python 3.7.13, the sklearn 1.0.2 package for ML methods,
and Keras 2.8.0 to construct neural networks. The simulation settings utilized in this investigation are
displayed in Table 5.

Table 5. Simulation parameters

Parameter Name Considered value
Hidden Units for LSTM [16, 32, 64]
Filters for CNN [16, 32, 64]
Memory Units for LSTM 32
Embedding Dimension 32
Hidden Units for FCN 250
Training batch 50
Optimizer Adam
Loss Function Binary Cross Entropy
Cross Validation 10-Fold

According to this experiment, we have considered 32 memory units for LSTM model and the filter
size for CNN is also considered as 32. However, these parameters were tested for a range of [16, 2, 64]. This
model uses a Fully connected layer and ReLU activation function is also applied for output layer. The
performance of this approach is measured with the help of confusion matrix which uses “Normal” and
“Attack” classes to identify the true positive, false positive, true negative and false negative. Table 6 shows
the obtained confusion matrix.
‒ True positive (TP): When a classifier properly predicts the positive class from the provided test set, it is
said to be true positive.
‒ True negative (TN): it demonstrates that, given the test data, the classifier model accurately forecasts the
negative class. The accuracy of the classifier is indicated by the true positive and true negative numbers.
These categories need to correspond with the true positive and true negative values, though.
‒ False positive (FP): indicates that the positive class was predicted by the classifier model in error.
‒ False negative (FN): indicates erroneously predicting the negative class by the classifier.
With the help of this confusion matrix, we measure the performance in terms of average accuracy, precision,
recall, and F-Score. These parameters can be expressed as [25]:
𝑇𝑃+𝑇𝑁
𝐴𝑐𝑐 = (13)
𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁

Int J Artif Intell, Vol. 14, No. 1, February 2025: 337-349

Int J Artif Intell ISSN: 2252-8938  347
TP
Recall = (14)
TP+TN

TP
P= (15)
TP+FP

2∗P∗Sensitivity
F= (16)
P+Sensitivity
Table 6. Confusion matrix
Predicted class
Actual class
Normal Attack
Normal True positive False negative
Attack False positive True negative

4.2. Comparative analysis

With the help of aforementioned parameters, we measure the performance of proposed model and
compare the obtained outcome with existing methods. Sun et al. [21] presented deep feature extraction model
and implemented several classifiers. Table 7 shows the obtained performance. In order to achieve this
performance, we have considered the combined dataset as discussed before where the final dataset is grouped
into two classes as “attack” and “benign” query.

Table 7. Comparative analysis for ML and DL (class wise analysis)

Classifier Class Precision Recall F1-score Accuracy
KNN Attack 0.98 0.95 0.97 0.96
Benign 0.95 0.98 0.90
Average 0.97 0.97 0.98
RF Attack 0.98 0.99 0.99 0.98
Benign 0.99 0.98 0.98
Average 0.99 0.99 0.99
LSTM Attack 0.88 0.87 0.87 0.87
Benign 0.87 0.88 0.87
Average 0.87 0.87 0.87
Proposed Attack 0.99 0.99 0.99 0.99
SQL-CB-GuArd Benign 0.99 0.99 0.99
Average 0.99 0.99 0.99

According to this experiment, the existing methods i.e. KNN, RF, and LSTM methods have used
deep feature extraction model whereas proposed model uses combination of pre-processing along with GRU
and attention mechanism to enhance the pattern learning process. Therefore, the average accuracy has been
reported as 0.96, 0.98, 0.87, and 0.99 by using KNN, RF, LSTM, proposed SQL-CB-GuArd, respectively.
Similarly, Fang et al. [22] presented Bi-LSTM based DL approach for SQLIA detection. Table 8 shows the
comparative analysis as discussed in [22]. The complete dataset is divided into 70% training and remaining
30% is used for testing purpose.

Table 8. Comparative analysis for ML and DL (overall performance analysis)

Algorithm Accuracy (%) Precision (%) Recall (%) F1-Score (%)
Random forest 97.88 97.90 97.88 97.89
SVM 88.22 88.76 88.22 88.36
Logistic regression 71.02 73.49 71.02 70.57
MLP 90.99 91.49 90.99 91.15
Bi-LSTM 99.26 99.26 99.25 99.24
Proposed SQL-CB-GuArd 99.80 99.55 99.60 99.25

Further, we compared the performance with the traditional DL and supervised ML models to
measure the performance of proposed approach. Table 9 shows the obtained performance in terms of
accuracy, precision, recall, and F1-score. The experimental analysis shows that the traditional LSTM model
reported the less accuracy as 62.32% whereas proposed SQL-CB-GuArd has reported highest classification
accuracy as 99.20%.

SQL-CB-GuArd: a deep learning mechanism for Structured query language … (AsifIqbal Sirmulla)
348  ISSN: 2252-8938

Table 9. Comparative analysis for DL methods

Method Accuracy (%) Precision (%) Recall (%) F1-score (%)
Autoencoder [14] 94.00 95.00 90.00 92.00
LSTM [10] 62.32 66.23 65.16 64.23
KNN [10] 82.69 71.51 88.56 79.13
DT [10] 92.33 89.58 89.74 89.66
SQLNN [10] 96.16 97.28 92.23 94.68
Proposed SQL-CB-GuArd 99.20 99.35 99.40 99.10

5. CONCLUSION
This articles discussed the threats to web based systems where SQLIA are considered as one of the
most crucial threat to these systems. Therefore, maintaining security in these applications remains a
challenging task. Several methods have been introduced to deal with this issue where ML based automated
methods have gained attention from research community. However, ML based methods face several
challenges which need to be addressed. Therefore, researchers have adopted DL based methods. In this work,
we adopt the DL method to improve the SQLIA detection performance. The proposed model performs data
pre-processing, vectorization and hybrid DL model to obtain the improved performance. The pre-processing
phase includes data decoding, generalization and tokenization. Further, vectorization model is implemented
to obtain the final feature vector. Finally, CNN, Bi-LSTM, and GRU models are employed. Further, attention
mechanism is also incorporated to improve the feature learning performance.

REFERENCES
[1] A. A. R. Farea et al., “Injections attacks efficient and secure techniques based on bidirectional long short time memory model,”
Computers, Materials and Continua, vol. 76, no. 3, pp. 3605–3662, 2023, doi: 10.32604/cmc.2023.040121.
[2] A. Kumar, S. Dutta, and P. Pranav, “Analysis of SQL injection attacks in the cloud and in WEB applications,” Security and
Privacy, vol. 7, no. 3, 2024, doi: 10.1002/spy2.370.
[3] M. Qasaimeh, R. A. Hammour, M. B. Yassein, R. S. Al-Qassas, J. A. L. Torralbo, and D. Lizcano, “Advanced security testing
using a cyber-attack forecasting model: A case study of financial institutions,” Journal of Software: Evolution and Process, vol.
34, no. 11, 2022, doi: 10.1002/smr.2489.
[4] S. Morgan, “2019 official annual cybercrime report,” Cybersecurity Ventures, Herjavec Group, 2019.
[5] L. Brew, L. Drazovich, and S. Wetzel, “The impact of COVID-19 on the security and resilience of the maritime transportation
system,” Proceedings of the 2021 IEEE International Conference on Cyber Security and Resilience, CSR 2021, pp. 510–517,
2021, doi: 10.1109/CSR51186.2021.9527935.
[6] H. Tupsamudre, S. Jain, and S. Lodha, “PhishMatch: a layered approach for effective detection of phishing URLs,” arXiv-
Computer Science, pp. 1-30, 2021.
[7] A. Arshad, A. U. Rehman, S. Javaid, T. M. Ali, J. A. Sheikh, and M. Azeem, “A systematic literature review on phishing and
anti-phishing techniques,” Pakistan Journal of Engineering and Technology, PakJET, vol. 4, no. 1, pp. 163-168, 2021.
[8] S. Heister and K. Yuthas, “How blockchain and AI enable personal data privacy and support cybersecurity,” in Blockchain
Potential in AI, 2022. doi: 10.5772/intechopen.96999.
[9] A. Dixit, J. Quaglietta, K. Nathan, L. Dias, and D. Nguyen, “Cybersecurity: guiding principles and risk management advice for
healthcare boards, senior leaders and risk managers,” Healthcare quarterly, vol. 25, no. 4, pp. 35–40, 2023, doi:
10.12927/hcq.2023.27019.
[10] N. Khan, J. Abdullah, and A. S. Khan, “Defending malicious script attacks using machine learning classifiers,” Wireless
Communications and Mobile Computing, vol. 2017, 2017, doi: 10.1155/2017/5360472.
[11] OWASP, “OWASP top ten: Top 10 web application security risks,” OWASP, 2021. Accessed: Feb. 26, 2024. [Online]. Available:
https://owasp.org/www-project-top-ten/
[12] F. K. Alarfaj and N. A. Khan, “Enhancing the performance of SQL injection attack detection through probabilistic neural
networks,” Applied Sciences, vol. 13, no. 7, 2023, doi: 10.3390/app13074365.
[13] N. Khan, J. Abdullah, and A. S. Khan, “Towards vulnerability prevention model for web browser using interceptor approach,”
2015 9th International Conference on IT in Asia: Transforming Big Data into Knowledge, CITA 2015, 2015, doi:
10.1109/CITA.2015.7349842.
[14] V. Jain, M. S. Gaur, V. Laxmi, and M. Mosbah, “Detection of SQLite database vulnerabilities in android apps,” Information
Systems Security (ICISS 2016), Springer, Cham, pp. 521–531, 2016, doi: 10.1007/978-3-319-49806-5_31.
[15] M. McPhee, Mastering kali Linux for web penetration testing. Birmingham, United Kingdom: Packt Publishing Ltd, Jun. 2017.
[16] D. E. Nofal and A. A. Amer, “SQL injection attacks detection and prevention based on neuro—fuzzy technique,” Studies in Big
Data, vol. 77, pp. 93–112, 2021, doi: 10.1007/978-3-030-59338-4_6.
[17] N. M. Sheykhkanloo, “A learning-based neural network model for the detection and classification of SQL injection attacks,” Deep
Learning and Neural Networks, pp. 450–475, 2019, doi: 10.4018/978-1-7998-0414-7.ch026.
[18] A. M. Vartouni, S. Mehralian, M. Teshnehlab, and S. S. Kashi, “Auto-encoder LSTM methods for anomaly-based web
application firewall,” International Journal of Information and Communication Technology, vol. 11, no. 3, pp. 49–56, 2019.
[19] N. Thalji, A. Raza, M. S. Islam, N. A. Samee, and M. M. Jamjoom, “AE-Net: novel autoencoder-based deep features for SQL
injection attack detection,” IEEE Access, vol. 11, pp. 135507–135516, 2023, doi: 10.1109/ACCESS.2023.3337645.
[20] N. Gandhi, J. Patel, R. Sisodiya, N. Doshi, and S. Mishra, “A CNN-BiLSTM based approach for detection of SQL injection
attacks,” Proceedings of 2nd IEEE International Conference on Computational Intelligence and Knowledge Economy, ICCIKE
2021, pp. 378–383, 2021, doi: 10.1109/ICCIKE51210.2021.9410675.
[21] H. Sun, Y. Du, and Q. Li, “Deep learning-based detection technology for SQL injection research and implementation,” Applied
Sciences, vol. 13, no. 16, 2023, doi: 10.3390/app13169466.

Int J Artif Intell, Vol. 14, No. 1, February 2025: 337-349

Int J Artif Intell ISSN: 2252-8938  349

[22] Y. Fang, J. Peng, L. Liu, and C. Huang, “WOVSQLI: Detection of SQL injection behaviors using word vector and LSTM,” ACM
International Conference Proceeding Series, pp. 170–174, 2018, doi: 10.1145/3199478.3199503.
[23] M. Kherbache, K. Amroun, and D. Espes, “A new wrapper feature selection model for anomaly-based intrusion detection
systems,” International Journal of Security and Networks, vol. 17, no. 2, pp. 107–123, 2022, doi: 10.1504/IJSN.2022.123298.
[24] I. Jemal, M. A. Haddar, O. Cheikhrouhou, and A. Mahfoudhi, “Performance evaluation of convolutional neural network for web
security,” Computer Communications, vol. 175, pp. 58–67, 2021, doi: 10.1016/j.comcom.2021.04.029.
[25] F. Dass, M. Dass, C. Feresa, and M. Foozy, “A comparative study of SQL injection detection using machine learning approach,”
Applied Information Technology And Computer Science, vol. 3, no. 2, pp. 19–031, 2022, doi: 10.30880/aitcs.2022.03.02.002.

BIOGRAPHIES OF AUTHORS

AsifIqbal Sirmulla holds Masters Degree in Computer Science from Reva

University, India in 2020. He also recieved his Bachelors from VTU, India in 2013. He is
currently Research Scholar at Reva University, Bangalore. His research includes machine
learning, data mining, and natural language processing. He has published 5 papers in
international journals and conferences. He can be contacted at email: [email protected].

Dr. Prabhakar Manickam is presently working as Professor in School of

Computer Science & Engineering, REVA University – Bangalore with an aim of designing a
project for farmers in India to cultivate appropriate crop in appropriate time for better yield.
He has published about 15 patents and has obtained 5 patent grants in various fields including
internet of things, image processing, education, and security in WSNs. He was awarded
`Excellence in Research’, by Novel Research Academy, Puducherry, India. He has delivered
many guest lectures on Internet of Things in `Border Security Force Signal Training School
(BSF-STS) for Officers’ - Bangalore. He can be contacted at email: [email protected].

SQL-CB-GuArd: a deep learning mechanism for Structured query language … (AsifIqbal Sirmulla)