Second International Conference on Augmented Intelligence and Sustainable Systems (ICAISS 2023)

IEEE Xplore Part Number : CFP23CB2-ART ; ISBN : 979-8-3503-2579-9

An Analysis of AI-based SQL Injection (SQLi)

Attack Detection
2023 Second International Conference on Augmented Intelligence and Sustainable Systems (ICAISS) | 979-8-3503-2579-9/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICAISS58487.2023.10250505

B. Brindavathi1 Aravind Karrothu2 Chunduru Anilkumar3

Department of Cyber Security in Department of Information Department of Information
CSE Technology Technology
GMR Institute of Technology GMR Institute of Technology GMR Institute of Technology
Rajam, India Rajam, India Rajam, India
[email protected] [email protected] [email protected]

Abstract—The SQL injection attack is a highly perilous applications related threats as per analysis from 2017 to 2021
vulnerability in the digital realm, especially for web pages, as [14]. In 2017 the injection attacks were in first position in
recognized by the Open Web Application Security Project terms of severity and then comes to 3rd position i.e., Injection
(OWASP) ranking. It is a type of code injection attacks. This attacks still in 3rd position due to high risk, so there is a need
kind of attacks basically breaches the virtual portion of the to prevent those risks with effective cost.
databases. Many web applications accept to store the user’s
private information (e.g., login credentials, credit card and other
account details etc.,) in the database over the internet. The
detection of SQL injection attack is going to be a tough task for
everyone because of an attacker can deploy various new type of
SQL injections in day-to-day life. There are many ways to
create/detect SQLi attacks by using open-source tools such as
Netsparker, SQLMap, jSQL Injection, Burp Suite, BBQSQL,
Nessus, and etc., So, the researchers have required an innovative
and never-before seen mechanisms for detecting these types of
attacks. But still, there is a lack of knowledge for input Fig.1. WEB Application architecture
validation. So, various authors have decided to improve the SQLi
attack detection by using various artificial intellectual Injection flaws, including SQL, NoSQL, OS, and LDAP
mechanisms like machine learning (ML) and deep learning (DL). injections, manifest when untrusted data is incorporated into a
Both ML and DL algorithms are solved many types of SQL command or query sent to an interpreter. This enables
injection attacks by using various classification or regression malicious actors to exploit the interpreter and execute
techniques. This study will focus on various recent methodologies unintended commands or gain unauthorized access to data.
that how to detect SQLi attacks along with which mechanism will According to the OWASP report analysis, the SQL injection
provide better performance among all the existing working attack ranks as one of the most prevalent web-based attacks
models. and the third most significant web application security risk in
2021. This attacks mainly injects malicious SQL code into a
Keywords—Structured Query Language (SQL), SQL Injection web application, allowing the attacker to view or modify a
(SQLi) attacks, Deep Learning (DL), Machine Learning (ML) and
database. An SQL injection (SQLi) attack, if successful, can
Open Web Application Security Project (OWASP)
lead to the unauthorized retrieval of sensitive data, such as
passwords, credit card information, or personal user data. The
I. INTRODUCTION below figure has explained the difference between malicious
In digital world, many of the organizations have used web- and genuine queries how they entered in the application.
based applications for accessing the data over the Internet to
perform different types of online communications. In the
realm of web applications and websites, user-entered data
during transactions is typically stored in a database. These
databases, often relational in nature, are interacted with using
a language known as Structured Query Language (SQL). Any
web application accepts the user’s data and produces SQL
statements as output. The OWASP Top 10 is a widely
recognized document aimed at raising awareness among
developers and web application security professionals. It
serves as a standard reference for identifying and
understanding the most significant security risks that web
applications face. The OWSP team released every year top 10
threat as per their analysis and the latest version have released Fig. 2. Normal-Malicious web application scenario
on sept. 2021. The Figure 2 represents the severity of web

979-8-3503-2579-9/23/$31.00 ©2023 IEEE 31

Authorized licensed use limited to: UNIV OF DERBY. Downloaded on February 06,2025 at 20:47:20 UTC from IEEE Xplore. Restrictions apply.
 Second International Conference on Augmented Intelligence and Sustainable Systems (ICAISS 2023)
IEEE Xplore Part Number : CFP23CB2-ART ; ISBN : 979-8-3503-2579-9

A. How SQLi attacks works multiple SELECT string into a unified result, which is
The attacker can give the input userID value as SELECT * subsequently included as part of the HTTP response.
FROM USERS WHERE user_id = ‘ ‘ OR ‘1’= ‘1’--’; As per 2) Inferential SQLi (Blind SQLi): Unlike in-band SQLi,
OR function, if any one of the input statements is true then the this method may require an extension of timeframe for an
irrespective of the other input, the result output is always attacker to exploit. The type of SQLi attack that relies on
TRUE. Then, the server may consider the malicious query as boolean values or time-based results to determine success is
normal and executes that. As final result, the sensitive data of known as Blind SQL Injection. In Blind SQL Injection
all the users will be revealed to him by the server [15]. The attacks, the attacker doesn't receive direct responses from the
figure 3 depicts the complete scenario of SQLi attack. application that indicate whether the injection was successful
or not. Instead, they infer the results by analyzing the behavior
or timing of the application's responses.
a) Boolean-based SQLi: It is a type of inferential SQLi
technique where an SQL query is sent to the database. The
response from the application varies depends on the input
query and it produces either TRUE or FALSE outcome.
b) Time-based SQLi: It is another type of inferential
SQLi technique that involves for transmitting various query
statements to the database, causing it to delay its response for
Fig. 3. SQLi attack scenario
a specified duration (in seconds). The response time serves as
an indicator to the attacker, revealing the input query status
Through an SQL injection attack, an attacker possesses the with TRUE or FALSE outcome.
ability to bypass authentication measures, gain unauthorized
access to a database, manipulate data, and even delete c) Out-of-band SQLi:. It is SQLi happens when the
information. In certain instances, SQLi can be leveraged to attacker cannot work in the similar channel for launching the
execute commands on the operating system, thereby providing outbreak and obtaining results. This type of attack is relatively
attackers with the means to carry out more destructive actions uncommon, primarily because it relies on specific
within a network of diverse types. characterstics being enabled on the server of database utilized
by the respective web-application.
B. Types of SQLi attacks
C. Traditional SQL prevention Techniques
SQLi have classified into three variants namely, In-band
SQLi, Out-of-band SQLi and Inferential SQLi [16]. The Given the vast number of pages and applications on a
classification has depicted in figure 4. website, it is practically impossible to manually inspect each
one, especially when updates are frequent and prioritizing
user-friendliness. To ensure the security of our database within
the server, security analysts and research developers
recommend implementing several key measures. Various
conventional identification mechanisms have been proposed to
achieve this goal.
1) Continuous Scanning and Penetration Testing: It
entails the periodic utilization of automated
vulnerability scanning tools to identify and evaluate
any existing vulnerabilities in a system. Another
Fig. 4. Types of SQLi attacks practice, known as penetration testing or "pen
1) In-band SQLi (Classic SQLi):. In the realm of SQL testing," involves simulating an attack on a system to
Injection attacks, in-band SQL Injection stands out as the most identify and exploit vulnerabilities.
widespread and readily exploitable. This particular attack 2) Restrict Privileges: This practice involves
method involves the attacker utilizing the same implementing restrictions on the access and
communication channel for both launching the attack and permissions granted to users and applications in
obtaining the results. The two main variations of in-band SQL relation to the database and its various components.
Injection are Error-based SQLi and Union-based SQLi. 3) Use Query Parameters: By utilizing query
parameters, it becomes possible to separate user input
a) Error-based SQLi: The success of this approach from the SQL query, significantly increasing the
relies on extracting information about the database structure difficulty for an attacker to inject malicious code into
through error-messages generated by the server-side database. the query.
b) Union-based SQLi: In this method, the attacker 4) Instant Protection with Web Application Firewall
utilizes the UNION SQL statement to merge the outcomes of (WAF): It has the capability to identify and prevent
malicious traffic from reaching the web application.

979-8-3503-2579-9/23/$31.00 ©2023 IEEE 32

Authorized licensed use limited to: UNIV OF DERBY. Downloaded on February 06,2025 at 20:47:20 UTC from IEEE Xplore. Restrictions apply.
 Second International Conference on Augmented Intelligence and Sustainable Systems (ICAISS 2023)
IEEE Xplore Part Number : CFP23CB2-ART ; ISBN : 979-8-3503-2579-9

It can be configured to detect and block recognized This model proposed by Marasahdih et al., for predicting
SQL injection payloads, ensuring enhanced security injected SQL statements from the web portal. This
against such attacks. methodology consists of three stages, whereas in first stage,
There are many traditional techniques or tools avail for the input statement is converted into AST (Abstract Syntax
identifying or mitigating SQLi attacks. It is possible to Tree) form then it was converted into minimal Static Single
discover SQL injection vulnerabilities using automated Assignment form (MSSA). The MSSA form is again entered
scanners. Automated scanners can employ different techniques into extraction stage that includes path generation, filtering
like static analysis, dynamic analysis, Boolean-based and and vulnerability context rules. In final stage, the authors have
studied various ML/DL algorithms for predicting detection
error-based testing to detect SQL injection vulnerabilities. But
performance. Through this methodology the authors have
those scanners not applicable for detect all the types of SQLi
concluded LSTM is the best solution for detect the SQLi
vulnerabilities. Even there are some organizations fails to attacks [2]. In this work, the authors have explained a robust
detect these attacks because many organizations encounter semantic query-featured ensemble-based learning model for
challenges such as outdated code, limited resources for testing SQLi attack detection with six stages. The all six stages are
and implementing changes, lack of awareness about helpful for classify the injected queries from the dataset. Data
application security, and infrequent updates to their web acquisition and tokenization is the first stage for split the
applications, leading to vulnerabilities. For these, we need to statement into various individual tokens. Semantic feature
solve those attacks by using statistical and artificial extraction is the next stage and it was applied by CBOW
intellectual mechanisms such as machine learning, deep mechanism. In third stage, data imbalance problem was
learning and etc., This work is analyzing various ML/DL overcome by data resampling technique. Then feature
based techniques and how to perform those techniques against selection was applied by Principle Component Analysis (PCA)
SQLi attacks. The next section highlights of literature review technique. The fifth stage was normalization mechanism for
with recent and major impact mechanisms. Then the best three verify the overfit problem. The final stage is for classifying
methodologies have discussed based on their performance SQLi attack detection through ensemble-based learning.
along with respective results and discussions. The final section Through this model, the authors have concluded in which this
concluded with future scope. model have given 98% of accuracy [3].
There was another 2-stage mechanism for providing both
II. LITERATURE REVIEW feature extraction and train the model. This is a compressive
In this section, we are going to discuss various existing and hybrid framework using various ML algorithms. This
methodologies to predicting SQLi attacking files from the hybrid mechanism has given best accuracy among all the
given dataset along with each technique’s limitations, future existing works what they have taken for the research but this
scope and also explained which performance metrics they approach was only applicable for small dataset and not for
have taken for the implementation. Each technique has large [4]. The authors have done SQLi attack detection by
implemented with their own style to give best performance. In using Q-learning reinforcement learning agents. Based on
this state-of-art, one or two articles have taken from the Markov decision process, this model was rectified the
analysis basis and most of the recent articles developed a new problems. But this mechanism only for union-based SQLi type
technique for predicting SQLi attacking files from the dataset. and not for rest of the other type of SQLi attacks [5]. Kasim. O
SQLiA detection can be enhanced through various ML/DL has detected a 4-stage method for detecting malicious SQL
techniques whereas features based ML techniques, Anomaly queries. In the first stage, various mechanisms have applied
identification techniques, Natural Language Processing (NLP) for the data clean. Then the updated data was extracted by 22
techniques and model-based techniques. While ML/DL different features. In third stage various detection algorithms
techniques can be valuable in identifying SQLi vulnerabilities, have used and then finally they did active prevention mode for
it is crucial to remember that they should not be relied upon as all the queries raised in the web application. This proposed
the only means of defense. It is equally important to model has maintained above 96% of accuracy for detecting
implement additional security measures, including input SQLi attacks. Also, through this model the author has
validation, parameterized queries, and secure coding practices. achieved 92% of accuracy for classifying the type of SQLi [6].
These measures work in conjunction with ML techniques to In this work, the authors have introduced a novel detection
collectively mitigate the risk of SQLi attacks. technique based on Long Short‐Term Memory (LSTM) and
ASTs. First, they have taken the raw SQL query statements
The authors have introduced a new deep neural network-
and converted into ASTs. Then Convert AST into symbol
based SQLi detection model (SQLNN) that was designed by
representation through Depth First Search (DFS) method. By
using characteristics of SQL statements. The overall model
using word embedding technique those symbols were
consists three modules whereas pre-process of the data, train
converted to vector representation also into a matrix form.
the model and model evaluation. In pre-process, the authors
Finally, LSTM has used for experimentation on the updated
have used TF-IDF (Term Frequency – Inverse Document
matrix representation. Through this model the authors have
Frequency) algorithm for removal of redundant words and it
achieved more accuracy than the existing [7].
keeps required words in those statements. This model contains
ReLU function for optimizing traditional loss and dropout The authors have proposed a signature-based SQLi attack
method is for improving the performance of the model. After detection model with three stages namely, data extraction,
valuation, this model maintained above 96% of accuracy [1]. train the data and detection of malicious queries from the data.

979-8-3503-2579-9/23/$31.00 ©2023 IEEE 33

Authorized licensed use limited to: UNIV OF DERBY. Downloaded on February 06,2025 at 20:47:20 UTC from IEEE Xplore. Restrictions apply.
 Second International Conference on Augmented Intelligence and Sustainable Systems (ICAISS 2023)
IEEE Xplore Part Number : CFP23CB2-ART ; ISBN : 979-8-3503-2579-9

Through this mechanism they have identified high accuracy in for SQLi attack detection through Artificial Neural Network
terms of identifying effected SQLi queries from the web (ANN). This method consist of three stages, data pre-
applications requests through source IP addresses [8]. A new processing, feature extraction, and then train the model by
light-weight approach has identified by Chen et al., to prevent using MLP and LSTM. After the evaluation, the authors have
SQLi attacks. This mechanism explained through two stages concluded that MLP have given high accuracy 99.67% when
firstly, data extraction done by using word embedding and the compared with LSTM [12]. This is another work based on
second stage is to detect the SQLi attack performance by the Knuth-Morris-Pratt (KMP) string matching algorithm to detect
help of various DL models. In data extraction, all the HTTP SQLi attacks. It was used to compare user’s input SQL
requests decoded by the help of Word2vec and generate statement with existed pattern of the injected SQL statement
various individual characters through word embedding whether any malicious entered into web application or not.
mechanism. These characters were trained by the both the After identification of malicious patterns, then the authors
CNN and MLP models. After usage of classifiers, the authors have designed a parse tree with those patterns. Then KMP
have concluded that MLP is the best one for this task with matching algorithm have used for filter the updated patterns.
98.57% of accuracy and other side 98.25% of accuracy have In this way, the authors have checked the all the raw input
resulted by CNN. Also, they have identified CNN was taking statement and concluded which are from malicious and normal
double time for train the data when compared with MLP [9]. queries [13]. There are so many review articles have existed
The authors have suggested that to block malicious SQL for giving informative knowledge on SQLi through various
queries by using another ML algorithm was Support Vector ML and DL [14-16]
Machine (SVM). This mechanism was concluded that memory
consumption can reduced and also can get the best Most of the existing algorithms explained with various
performance through F-score [10]. ML/DL algorithms to identifying whether the user’s query is
malicious or genuine from the respective dataset. Each
Aliero et al., have recommended a component-based algorithm performed based on certain metrics as accuracy,
technique to reduce the frequency of false results. This precision, recall and F1-score and also, we have analyzed all
proposed methodology has compared with current open source the algorithms with their limitations and future scope. The
SQLi attack detection tools and identified some sort of complete state-of-art analysis done through a table shown in
limitations and then improved. Also, through this mechanism, below.
they can well perform in SQLi vulnerability prediction [11].
Tang et al., have presents an efficient and simple mechanism
Table 1. An analysis of various ML/DL algorithms

Refer. Year of Limitations / Areas for future Performance

Author Names Methodology Used
No. Publish Advantages research metrics

Solves Overfitting
[1] Zhang et al., 2022 SQLNN Model --- Accuracy is >96%
Problem effectively

The authors will work

Feature extraction- Only applicable for
[2] Marashdih, et al., 2022 out on rest of the Accuracy is 98.7%
based Model PHP source codes
languages.

semantic query-
Gowtham, M., et Only suitable for Will do on XML Accuracy is 98%.
[3] 2022 featured ensemble
al., SQL queries intrusion detection and F1 score
learning model

Only applicable for

Comprehensive
small dataset and Will do for large
[4] Deriba, F., et al., 2022 Framework and Accuracy is 98%
takes more time for dataset
hybrid model
execution

Only applicable for

Markov decision Will do on another set
[5] Erdődi, L., et al., 2021 Union based SQLi Time complexity
process of SQLi attacks
attacks

Can detect on three

[6] Kasim, Ö. 2021 Decision tree model various SQLi ---- Accuracy is > 98%
attacks

unusable in
Deep LSTM NN on Reduction of FPR w.r.t
[7] Zhuo, Z., et al., 2021 some sorts of TPR, FPR and FNR
ASTs TP
realistic contexts

979-8-3503-2579-9/23/$31.00 ©2023 IEEE 34

Authorized licensed use limited to: UNIV OF DERBY. Downloaded on February 06,2025 at 20:47:20 UTC from IEEE Xplore. Restrictions apply.
 Second International Conference on Augmented Intelligence and Sustainable Systems (ICAISS 2023)
IEEE Xplore Part Number : CFP23CB2-ART ; ISBN : 979-8-3503-2579-9

Signature based Will do for large

Azman, M. A., et Limited amount of
[8] 2021 detection with KB datasets and improve Accuracy is 93%
al., dataset used
features accuracy

focus on advanced Accuracy of MLP is

SQLIA model based Not focused on all
SQL injection like 98.57%
[9] Chen, D., et al., 2021 on ANN [MLP- types of SQLi
second-order and Accuracy of CNN is
CNN] attacks
hybrid SQLi 98.25%

To take advantage of
Support Vector
Q-learning is the query workload
Meduri, V. V., et Machine (SVM)
[10] 2021 best performer in prediction for the F-1 Score
al., with machine
the implementation purpose of
learning
reorganizing data.

It could perform Will enhance for the F Measure,

Aliero, M. S., et Component based
[11]
al.,
2020 on all type of second-order SQLi Precision and
technique
SQLi attacks attack detection Recall
Accuracy of MLP is
SQLIA model based
LSTM will give Will do with other 99.92%
[12] Tang, P., et al., 2020 on ANN [MLP-
best accuracy advanced NN models Accuracy of LSTM
LSTM]
is 98.69%

It can block the

Knuth-Morris-Pratt They are working with
Abikoye, O. C., et system based on
[13] 2020 (KMP) string other set of scripting ----
al., MAC address if it is
matching algorithm languages
SQLIA based query

III. CONCLUSION [5] Erdődi, L., Sommervoll, Å. Å., & Zennaro, F. M. (2021). Simulating
SQL injection vulnerability exploitation using Q-learning reinforcement
According to the OWASP report, SQLIA is a noteworthy learning agents. Journal of Information Security and Applications, 61,
threat to online services, potentially leading to severe privacy 102903.
along with security implications. Various ML/DL algorithms [6] Kasim, Ö. (2021). An ensemble classification-based approach to detect
have demonstrated notable accomplishment in identifying attack level of SQL injections. Journal of Information Security and
various types of online attacks. We have analyzed around Applications, 59, 102852.
sixteen articles based on ML/DL models that explored [7] Zhuo, Z., Cai, T., Zhang, X., & Lv, F. (2021). Long short‐term memory
on abstract syntax tree for SQL injection detection. IET Software, 15(2),
research on SQL injection attacks and the application of 188-197.
ML/DL techniques. Among the identified techniques, neural [8] Azman, M. A., Marhusin, M. F., & Sulaiman, R. (2021). Machine
network-based approaches were found to be commonly used learning-based technique to detect SQL injection attack. Journal of
for detecting various forms of SQL injection attacks. Computer Science.
Furthermore, the majority of DL models exhibited superior [9] Chen, D., Yan, Q., Wu, C., & Zhao, J. (2021). Sql injection attack
detection and prevention techniques using deep learning. In Journal of
accuracy in identifying SQLi attacks. Notably, only a few Physics: Conference Series (Vol. 1757, No. 1, p. 012055). IOP
studies focused on employing ML for generating adversarial Publishing.
SQLi attack queries. Based on our analysis, we determined [10] Meduri, V. V., Chowdhury, K., & Sarwat, M. (2021). Evaluation of
that the MLP classifier emerged as the best performing model machine learning algorithms in predicting the next SQL query from the
in terms of accuracy. In future research, we intend to explore a future. ACM Transactions on Database Systems (TODS), 46(1), 1-46.
broader range of ML/DL models for both generating and [11] Aliero, M. S., Qureshi, K. N., Pasha, M. F., Ahmad, A., & Jeon, G.
(2020). Detection of structure query language injection vulnerability in
detecting SQLi attacks. We are trying design an appropriate web driven database application. Concurrency and Computation:
DL model for detecting SQLi attack in e-platforms. Practice and Experience, 34(13), e5936.
[12] Tang, P., Qiu, W., Huang, Z., Lian, H., & Liu, G. (2020). Detection of
References SQL injection based on artificial neural network. Knowledge-Based
Systems, 190, 105528.
[1] Zhang, W., Li, Y., Li, X., Shao, M., Mi, Y., Zhang, H., & Zhi, G.
(2022). Deep neural network-based SQL injection detection method. [13] Abikoye, O. C., Abubakar, A., Dokoro, A. H., Akande, O. N., &
Security and Communication Networks, 2022. Kayode, A. A. (2020). A novel technique to prevent SQL injection and
cross-site scripting attacks using Knuth-Morris-Pratt string match
[2] Marashdih, A. W., Zaaba, Z. F., & Suwais, K. (2022). Predicting input algorithm. EURASIP Journal on Information Security, 2020(1), 1-14.
validation vulnerabilities based on minimal SSA features and machine
learning. Journal of King Saud University-Computer and Information [14] Kritikos, K., Magoutis, K., Papoutsakis, M., & Ioannidis, S. (2019). A
Sciences, 34(10), 9311-9331. survey on vulnerability assessment tools and databases for cloud-based
web applications. Array, 3, 100011.
[3] Gowtham, M., & Pramod, H. B. (2022). Semantic query-featured
ensemble learning model for SQL-injection attack detection in IoT- [15] Alghawazi, M., Alghazzawi, D., & Alarifi, S. (2022). Detection of sql
ecosystems. IEEE Transactions on Reliability, 71(2), 1057-1074. injection attack using machine learning techniques: A systematic
literature review. Journal of Cybersecurity and Privacy, 2(4), 764-777.
[4] Deriba, F., Salau, A. O., Mohammed, S. H., Kassa, T. M., & Demilie,
W. B. (2022). Development of a compressive framework using machine [16] Alsalamah, M., Alwabli, H., Alqwifli, H., & Ibrahim, D. M. (2021). A
learning approaches for SQL injection attacks. PRZEGLĄD Review Study On Sql Injection Attacks, Prevention, And Detection.
ELEKTROTECHNICZNY, 1(7), 183-189.

979-8-3503-2579-9/23/$31.00 ©2023 IEEE 35

Authorized licensed use limited to: UNIV OF DERBY. Downloaded on February 06,2025 at 20:47:20 UTC from IEEE Xplore. Restrictions apply.