Detection of Structured Query Language Injection Attacks Using Machine Learning Techniques

International Journal of Computer Science & Information Technology (IJCSIT) Vol 15, No 4, August 2023
DOI:10.5121/ijcsit.2023.15402 13
DETECTION OF STRUCTURED QUERY LANGUAGE
INJECTION ATTACKS USING
MACHINE LEARNING TECHNIQUES
Taapopi John Angula and Valerianus Hashiyana
Department of Computing, Mathematical and Statistical Sciences, University of
Namibia, Windhoek, Namibia
ABSTRACT
This paper presents a comparative analysis of various machine learning classification models for
structured query language injection prevention. The objective is to identify the best-performing model in
terms of accuracy on a given dataset. The study utilizes popular classifiers such as Logistic Regression,
Naive Bayes, Decision Tree, Random Forest, K-Nearest Neighbors, and Support Vector Machine. Based on
the tests used to evaluate the performance of the classifiers, the Naïve Bayes gets the highest level of
accurate detection. The results show a 97.06% detection rate for the Naïve Bayes, followed by
LogisticRegression (0.9610), Support Vector Machine (0.9586), RandomForest (0.9530), DecisionTree
(0.9069), and K-Nearest Neighbor (0.6937). The code snippet provided demonstrates the implementation
and evaluation of these models.
KEYWORDS
Classification models, SQL-I, Python, Machine learning, Evaluations
1. INTRODUCTION
SQL (structured query language) injection vulnerabilities encompass the measures inculcated to
protect individual websites, network systems, and database infrastructures that host different
applications. As viewed by [1], they cause sections of institution’s web pages to hang
significantly, further increasing the spread of viruses, network paralysis, data privacy breaches,
and remote control of the servers. According to [2], the straightforward nature of SQL injection
vulnerabilities makes them a prime choice for network attacks aimed at infiltrating targeted
systems. One of the common ways hackers intrude into databases is through an SQL injection
attack [2]. This fact explains why programmers rely on using different modes to create codes that
may be used in addressing these vulnerabilities [3]. However, several challenges have been
exhibited that span from the limitations in accessing input data’s authenticity during the design
and the development of codes, which poses security concerns [4]. To this effect, SQL-I is a web
application security vulnerability that permits an attacker to execute mischievous SQL statements
on a web application’s backend database hosted on an SQL server.
This paper focuses on SQL injection attack detection and prevention that specifically target user
input fields in web server applications, such as XAMP. The objective of these attacks is to exploit
vulnerabilities in the system by injecting malicious SQL statements. SQL injection is a type of
web attack where an attacker inserts inputs that are executed as sequential queries. The targeted
system is typically unprepared to handle such input, leading to potential data disclosure and

14
unauthorized access. The consequences of a successful attack can be severe, impacting the
security aspects of confidentiality, integrity, and data availability.
During an SQL injection, the malicious actor inserts an SQL statement into an input field on a
client web application i.e., login page username field to access a database server. SQL is used to
represent queries to database management systems (DBMSs). Maliciously injected SQL statement
is designed to extract or modify data from the database server [5].
A successful injection can lead to verification and circumvention, as well as alterations to the
database through the insertion, alteration, or removal of data, resulting in data loss and potential
destruction of the entire database. Additionally, this type of attack has the capability to
overwhelm and execute commands on DBMS, often leading to more severe consequences [5].
Consequently, SQL injection attacks pose significant risks to organizations. Numerous studies
have been conducted to address this threat, introducing diverse artificial intelligence (AI)
approaches for detecting SQL injection attacks using machine learning and deep learning models
[5]. This approach includes threat detection by learning from past data that represents both attacks
and normal data.
On that bases historical data has become valuable for teaching machine learning (ML) models to
recognize attack patterns, comprehend detected injections, and even predict future attacks before
they occur [5]. In light of this the malicious actor and security defender of SQL injection attacks
must possess an understanding of how the SQL language operates in order to exploit its
vulnerabilities [6]. To extract or modify data from a database, queries must be written using the
SQL language and adhere to a standardized syntax, such as: "SELECT * FROM users WHERE
username = 'John_Doe'".
The above query is employed to fetch data from a table named "users." In this query, the asterisk
(*) denotes all columns, while the condition "WHERE username = 'John_Doe'" specifies that only
rows with the username 'John_Doe' should be selected. This query allows user to retrieve user
records from the database that match the specified username ‘John_Doe’.
Now, let’s assume a malicious actor wants to inject malicious SQL to the given query; i.e.,
" SELECT∗ FROM users WHERE username = ′John_Doe′ "
What would most likely happen is that the malicious actor would add ' OR '1'='1 to the end of the
above query. This would result in all the details in the table called users to be returned as a result
of the that addition. The typical representation would be type like this:
"SELECT ∗ FROM users WHERE username = ′John_Doe′OR′1′ = ′1′ "
All the data from the users table will be returned because the added string '1'='1' always equals
true and adding OR informs the database that it should return all the data from the users table
WHERE the username = ‘John_Doe’ or 1 = 1. Meaning this query would always return all the
other users’ details because 1 is aways equal to 1. This type of attack is sometimes referred to as a
tautology type attack [7]. In addition, tautology type attacks aren’t the only type of attacks that
represent a threat to databases and web applications. According to [5], alternative techniques,
apart from tautology, can be employed, like when a malevolent intruder purposely inserts a wrong
query to force the database server into producing a typical error page. This page could potentially
contain valuable insights that could assist the intruder in comprehending the database, thereby
facilitating a more sophisticated assault. The term "UNION" can similarly be leveraged to extract
data, in addition to numerous other approaches rooted in the same concept of exploiting SQL

15
syntax to extract or modify information within the designated database. The below example
shows this other method:
SELECT * FROM users WHERE username=''
UNION
SELECT * FROM users WHERE 1=1;
Given the dangers identified from SQL injections, the question that remains is how an
organization or developer can detect and prevent such attacks from affecting them?
As indicated earlier one method could be using ML, where a classifier is trained using a large
dataset of labeled data then allowed to identify if a username contains any SQL injection or not.
The outcome of which is to identify and reject requests to the database if suspicious activity is
identified. The labeled dataset is used on a ML that employs a supervised learning method. This
translates to bad data being labeled as bad and good data as good, this helps the ML learn and
recognize the difference between good and bad data such that when it come cross any of the two
it would be able to correctly assume and act.
[5], further elaborate on supervised learning, which involves using a labeled training dataset to
train a classifier. In this approach, the input data is already labeled as normal or abnormal,
enabling the algorithm to establish a straightforward mapping between the input and the known
output. The algorithm iteratively modifies its weights until achieving an acceptable classification
accuracy. Subsequently, a separate test dataset is employed to evaluate the classifier's
performance. If the results fall within an acceptable accuracy range, the classifier becomes
capable of detecting novel data that was not part of the training or testing process. However,
generating and labeling the training and testing data can be time-consuming, especially for
complex attacks. Supervised learning encompasses two main categories: classification and
regression algorithms. Popular examples of supervised learning algorithms include Bayesian
networks, decision trees, support vector machines, K-nearest neighbors, and neural networks. The
foremost input of this paper is to provide a systematic review of machine learning techniques and
input validation that are used to detect and prevent injections. Through this comprehensive
review, our objective is to provide the insights and enhance the comprehension of the
convergence of an SQL injection attack and artificial intelligence, thereby keeping researchers
well-informed of the current developments in this field.
2. OBJECTIVES
The main objective of this research is to detect structured query language injection attacks (SQL-
I) using machine learning techniques and achieve the following sub-objectives:
a) Evaluate the effectiveness of different machine learning classification models for
SQLdetection.
b) Identify the model(s) that demonstrate the highest accuracy and performance in detecting
SQL-I attacks.
c) Determine the most suitable model(s) for building a robust SQL-I detection system.
d) Compare the models based on metrics such as accuracy, precision, recall, F1-score, and
computational efficiency.
e) Provide insights and recommendations for selecting and implementing machine learning
models for SQL-I prevention in real-world applications.

16
3. BACKGROUND AND RELATED WORK
[8] provide a foundational review of the types of SQL injection attacks being broadly categorized
into In-band and Out-of-band. The former occurs when the attacker can receive the results of the
injected SQL query through the same channel as the web application. On the other hand, out-
ofband attacks involve the attacker receiving the results through a different channel, such as email
or a separate website. They further suggest some counter measures such as input validation,
parameterized queries, and stored procedures. Input validation involves checking the user input
for malicious SQL code before it is sent to the database. While, parameterized queries and stored
procedures use pre-defined SQL commands that do not allow for arbitrary user input.
[9] add to the foundation laid by stating that on top of these counter measures, machine learning
techniques can be applied to detect SQL injection attacks.These methods include decision trees,
neural networks, and support vector machines. By training the machine learning models with a
large dataset of legitimate and malicious queries, the models can detect the malicious queries with
high accuracy.
In conclusion, the combination of both methods provides an efficient way to detect and prevent
SQL injection attacks.
4. METHODOLOGY
The research adopts a comparative analysis methodology to assess and identify machine learning
classification models that are most effective in preventing SQL injection attacks on a given
dataset. The primary objective is to evaluate the performance of different models and determine
the optimal model(s) for SQL injection prevention.
4.1. Research Design
The research employed a quantitative research design to conduct a comparative analysis of
machine learning classification models for SQL injection prevention on a given dataset. The study
encompassed a review of existing research materials, including journal articles, conference
papers, articles, blogs, books, webpages, and academic tools.The research project was divided
into several stages to ensure a systematic and comprehensive approach. The stages followed in the
study are as follows:
4.1.1. Dataset Selection
4.1.1.1. Identified a suitable dataset specifically designed for evaluating SQL injection
prevention models.
4.1.1.2. Ensured the dataset encompassed a diverse range of SQL injection attack instances
and benign queries, representing real-world scenarios.
4.1.2. Preprocessing and Feature Extraction
4.1.2.1. Performed necessary preprocessing steps on the dataset, such as data cleaning,
handling missing values, and ensuring data quality.
4.1.2.2. Extracted meaningful features from the dataset using appropriate techniques, such as
tokenization, vectorization, or n-gram representation, to represent the SQL queries
effectively.

17
4.1.3. Model Selection
4.1.3.1. Evaluated various machine learning classification models suitable for SQL injection
prevention, such as decision trees, logistic regression, support vector machines,
neural networks, and others.
4.1.3.2. Considered factors such as model performance, interpretability, computational
requirements, and scalability to select the most promising models for further
evaluation.
4.1.4. Model Training and Evaluation
4.1.4.1. Trained the selected models using the preprocessed dataset, allowing the models to
learn the patterns and characteristics of SQL injection attacks and normal queries.
4.1.4.2. Evaluated the trained models' performance using appropriate metrics, including
accuracy, precision, recall, and F1-score, to assess their effectiveness in SQL
injection prevention.
4.1.5. Comparative Analysis
4.1.5.1. Conducted a comprehensive comparative analysis of the trained models, considering
their strengths, weaknesses, and suitability for SQL injection prevention.
4.1.5.2. Compared the models based on metrics, interpretability, and scalability to identify
the optimal model(s) for SQL injection prevention on the given dataset.
4.1.6. Results and Conclusions
4.1.6.1. Analyzed and interpreted the results obtained from the comparative analysis.
4.1.6.2. Drew conclusions on the performance and effectiveness of the different machine
learning classification models in SQL injection prevention.
4.1.6.3. Provided recommendations and insights based on the findings to guide future
research and implementation of SQL injection prevention systems.
By following these stages, the research project aimed to provide valuable insights into selecting
the most suitable machine learning classification models for SQL injection prevention,
contributing to the development of effective security measures for database systems and web
applications.
4.2. Dataset Selection
To conduct the study, A dataset containing both malicious and benign payloads was needed.
Specifically the SQL statements for SQL injection (SQL-i). Our datasets primarily focused on
payloads transmitted via HTTP requests to web applications. It's important to note that our
research did not cover other types of XSS attacks, such as "reflected." The collection obtained
encompasses a wide range of payloads, including obfuscated and non-obfuscated scripts, as well
as both long and short scripts, covering SQL-i.
Table 1 displays all the harmful and benign scripts collected from this data, encompassing SQL
injections, and regular text. It's worth noting that the datasets contain a limited number of SQL
injections, which is significant given the presence of malicious payloads within financial
institutions' systems. Web pages have a short lifespan in terms of appearance, making datasets
containing such attack payloads scarce.

18
Table 1. Benign and Malicious Scrips
Type Benign Malicious Total
SQL-i 1,050 3,150 4,200
The acquired data was used to construct datasets with the purpose of providing a simulation of
intrusions in Namibia’s banking system. The dataset was split into two parts: one portion was
used for training the models, while the other half was used for evaluating their performance. Once
the classifiers were developed, the testing datasets were used in the evaluation process.
4.3. Preprocessing and Feature Extraction
Jupiter Notebook environment for python 3 was used in the entirety of this study. This
environment was best suited for this study because it comes preinstalled with the vast majority of
data analysis tools and machine learning libraries.
The pandas libraries was used to conduct the data analysis and manipulation via the following
command:
𝑖𝑚𝑝𝑜𝑟𝑡𝑝𝑎𝑛𝑑𝑎𝑠 𝑎𝑠 𝑝𝑑 # 𝐼𝑚𝑝𝑜𝑟𝑡𝑡ℎ𝑒 𝑝𝑎𝑛𝑑𝑎𝑠 𝑙𝑖𝑏𝑟𝑎𝑟𝑦
𝑓𝑜𝑟 𝑑𝑎𝑡𝑎 𝑚𝑎𝑛𝑖𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛 𝑎𝑛𝑑 𝑎𝑛𝑎𝑙𝑦𝑠𝑖𝑠
In addition, the large dataset was imported into Jupyter Notebook as a Pandas dataframe using:
𝑝𝑑. 𝑟𝑒𝑎𝑑_𝑐𝑠𝑣command. The actual implementation is:
𝑑𝑓 = 𝑝𝑑. 𝑟𝑒𝑎𝑑_𝑐𝑠𝑣(′𝑑𝑎𝑡𝑎𝑠𝑒𝑡/𝑠𝑞𝑙𝑖. 𝑐𝑠𝑣′, 𝑒𝑛𝑐𝑜𝑑𝑖𝑛𝑔 = ′𝑢𝑡𝑓 − 16′)
The above simply reads the csv file into a pandasdataFrame, specify the file path and encode as
'utf-16' since the file was encoded in UTF-16 format.
Given the data frame being defined as df, a standard way to preview the it is by using the ℎ𝑒𝑎()
command. This command comes after the variable name of the dataframe, which in this case was
df. Thus, to preview the dataframe command is: 𝑑𝑓. ℎ𝑒𝑎() this resulted in the first first rows of
the data frame being displayed:
Figure 1. Dataframe output

19
Also, the following command 𝑑𝑓. 𝑑𝑒𝑠𝑐𝑟𝑖𝑏() was used to generate a summary of the dataframe's
mean, standard deviation and interquartile range:
Figure 2. Dataframe summary
Furthermore, to fully understand the dataframe an inspection of the number of missing values per
variable was needed and this was conducted using the 𝑑𝑓. 𝑖𝑠𝑛𝑢𝑙𝑙(). 𝑠𝑢() command. This
command output the following:
Figure 3. Inspecting missing values using isnull().sum()
The above indicated that there are 13 missing values in the column Sentence thus the need to use
the 𝑑𝑓. 𝑑𝑟𝑜𝑝𝑛() command to delete the rows with no values was valuable but consequently
resulted in 13 few rows from our initial 4200 rows.
Equally important the need to perform split validation from the dataset was paramount. This step
involved importing the necessary libraries through the following command:
𝑓𝑟𝑜𝑚 𝑠𝑘𝑙𝑒𝑎𝑟𝑛. 𝑚𝑜𝑑𝑒𝑙_𝑠𝑒𝑙𝑒𝑐𝑡𝑖𝑜𝑛 𝑖𝑚𝑝𝑜𝑟𝑡𝑡𝑟𝑎𝑖𝑛_𝑡𝑒𝑠𝑡_𝑠𝑝𝑙𝑖𝑡
This command simply imports train_test_split function for splitting data into training and testing
sets.
This step was implemented using the following lines of code:
𝑋 = 𝑑[′𝑆𝑒𝑛𝑡𝑒𝑛𝑐𝑒′] , in essence is assigning the 'Sentence' column as the feature (X) variable and
𝑦 = 𝑑𝑓[′𝐿𝑎𝑏𝑒𝑙′], as the target (y) variable. Furthermore, given the two values, X = the input

20
features and y = the output labels, creating the needed training and test data was optimal. Using
the standard split validation parameters figure seen below, the split validation was created.
Table 2. Split validation parameters
Parameter Argument Explanation Default
test_size Float(between 0 and
1.0)
Proportion of test size, i.e.
0,3 = 30%
0.25 (if train_size is not
specified)
Integer The number of test
samples, i.e, 40 = for 40
test samples
None Test value is automatically
set to complement that of
train_size.
train_size
(optional)
Float (between 0 and
1.0)
Proportion of test size, i.e.
0,7 # 70%
Integer Number of test samples,
i.e. 60 = for 60 test
samples
random_state
(optional)
Integer A seed number
(integer) that can be reused
to replicate the same
random split. This ensures
the model uses the same
data split each time it's
used.
None A random seed number is
used.
None
Shuffle
(optional)
TRUE Data is shuffled TRUE
FALSE Data is not shuffled
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.3, random_state = 10
,shuffle=True)
The above validation split consists of the following; X: The input features, y: The output labels,
test_size: The proportion of the dataset to include in the testing set. Here, it is set to 0.3, which
means 30% of the data will be used for testing, random_state: The random state used for shuffling
the data before splitting. It ensures reproducibility of the split. In this case, it is set to 10 and
lastly, shuffle: If True, the data will be shuffled before splitting. This helps in randomizing the
data, which can be useful to avoid any bias in the splitting process.
With the validation completed, the data needed to be converted into numerical representation
based on the frequency of words (or n-grams) in the text using the CountVectorizerclasss. This
class was imported using the following command:
𝑓𝑟𝑜𝑚 sklearn. feature_extraction. text import CountVectorizer
Secondly an instance of CountVectorizer is created using:
𝑣𝑒𝑐𝑡𝑜𝑟𝑖𝑧𝑒𝑟 = 𝐶𝑜𝑢𝑛𝑡𝑉𝑒𝑐𝑡𝑜𝑟𝑖𝑧𝑒𝑟()

21
Then, the fit_transform method is applied to the training data X_train. This step fits the vectorizer
to the training data and transforms the text into a numerical representation and thereafter, the
result is stored in X_train_transformed.
𝑋_𝑡𝑟𝑎𝑖𝑛_𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑑 = 𝑣𝑒𝑐𝑡𝑜𝑟𝑖𝑧𝑒𝑟. 𝑓𝑖𝑡_𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚(𝑋_𝑡𝑟𝑎𝑖𝑛)
Finally, the transform method is applied to the testing data X_test using the already fitted
vectorizer. This step transforms the testing data into the same numerical representation as the
training data. The result is stored in X_test_transformed.
𝑋_𝑡𝑒𝑠𝑡_𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑑 = 𝑣𝑒𝑐𝑡𝑜𝑟𝑖𝑧𝑒𝑟. 𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚(𝑋_𝑡𝑒𝑠𝑡)
After this transformation, one can use the transformed data X_train_transformed and
X_test_transformed for training and evaluating machine learning models.
4.4. Model Selection
In order to select an appropriate machine learning model, a best suited model needed to be
identified using training date and thus allow for a best accuracy to be utilized as performance
measure and finally the actual model to be used.
The first step was to import classification models into the workbook. Classification models were
used because the objective is to build a SQL-I detection system that would receive a query and
determine if that query is normal SQL or SQL-I. The below models were imported:
𝑓𝑟𝑜𝑚 𝑠𝑘𝑙𝑒𝑎𝑟𝑛. 𝑙𝑖𝑛𝑒𝑎𝑟_𝑚𝑜𝑑𝑒𝑙𝑖𝑚𝑝𝑜𝑟𝑡𝐿𝑜𝑔𝑖𝑠𝑡𝑖𝑐𝑅𝑒𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛
𝑓𝑟𝑜𝑚 𝑠𝑘𝑙𝑒𝑎𝑟𝑛. 𝑛𝑎𝑖𝑣𝑒_𝑏𝑎𝑦𝑒𝑠 𝑖𝑚𝑝𝑜𝑟𝑡𝑀𝑢𝑙𝑡𝑖𝑛𝑜𝑚𝑖𝑎𝑙𝑁𝐵
𝑓𝑟𝑜𝑚 𝑠𝑘𝑙𝑒𝑎𝑟𝑛. 𝑡𝑟𝑒𝑒 𝑖𝑚𝑝𝑜𝑟𝑡𝐷𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑇𝑟𝑒𝑒𝐶𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑟
𝑓𝑟𝑜𝑚 𝑠𝑘𝑙𝑒𝑎𝑟𝑛. 𝑒𝑛𝑠𝑒𝑚𝑏𝑙𝑒 𝑖𝑚𝑝𝑜𝑟𝑡𝑅𝑎𝑛𝑑𝑜𝑚𝐹𝑜𝑟𝑒𝑠𝑡𝐶𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑟
𝑓𝑟𝑜𝑚 𝑠𝑘𝑙𝑒𝑎𝑟𝑛. 𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑟𝑠 𝑖𝑚𝑝𝑜𝑟𝑡𝐾𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑟𝑠𝐶𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑒𝑟
𝑓𝑟𝑜𝑚 𝑠𝑘𝑙𝑒𝑎𝑟𝑛. 𝑠𝑣𝑚 𝑖𝑚𝑝𝑜𝑟𝑡𝑆𝑉𝐶
Secondly, an array initializing the models was implemented:
models = [ LogisticRegression(),
MultinomialNB(),DecisionTreeClassifier(),RandomForestClassifier(),KNeighborsClassifier(),S
VC() ]
Thirdly, a variable called best_model is initialized to None, and best_accuracy is set to 0.0
initially. A loop is used to calculate the accuracy for each model, compare it with the current
best_accuracy value and if the new accuracy is higher, it updates best_accuracy and assign the
corresponding model to best_model. Once the loop completes, the results are printed out showing
the name of the best model along with its accuracy.
By using this method of model selection one can identify the model that achieved the highest
accuracy on the dataset. The code below indicates how this was done:
best_model = None
best_accuracy = 0.0
for model in models:

22
model.fit(X_train_transformed, y_train) y_pred = model.predict(X_test_transformed)
accuracy = accuracy_score(y_test, y_pred)
print(f"{model.__class__.__name__} accuracy: {accuracy}")
if accuracy >best_accuracy: best_accuracy = accuracy
best_model = model
print (f"nThe best model is: {best_model.__class__.__name__} with an accuracy of
{best_accuracy}")
4.5. Model Training and Evaluation
The results of the above model selection indicated that the best model for the dataset used was the
MultinomialNB which is the Naive Bayes classifier for multinomial models. Below are the results
of the model selection:
Figure 4. Results of the model selection
The above figure indicates that the Naïve Bayes classifier was best suited for the dataset with an
accuracy of 97 %.
Given this high accuracy rate, the MultinomialNB could then be set as the main model to train
with the training data. This was done by initializing the model variable to MultinomialNB; 𝑚𝑜𝑑𝑒𝑙
= 𝑀𝑢𝑙𝑡𝑖𝑛𝑜𝑚𝑖𝑎𝑙𝑁(). Once the model was set, model needed to be fitted to the training data
i.e., fitting the model to the input features X_train_transformed and the corresponding target
labels y_train.;
𝑚𝑜𝑑𝑒𝑙. 𝑓𝑖𝑡(𝑋_𝑡𝑟𝑎𝑖𝑛_𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑑, 𝑦_𝑡𝑟𝑎𝑖𝑛)
In addition, once the model was fitted, it allowed for an evaluation to be conducted by making
predictions on the test data using the trained model. This is carried out by taking the input features
X_test_transformed and returns the predicted labels for the corresponding test samples;
𝑚𝑜𝑑𝑒𝑙_𝑝𝑟𝑒𝑑𝑖𝑐𝑡= 𝑚𝑜𝑑𝑒𝑙. 𝑝𝑟𝑒𝑑𝑖𝑐𝑡(𝑋_𝑡𝑒𝑠𝑡_𝑡𝑟𝑎𝑛𝑠𝑓𝑜𝑟𝑚𝑒𝑑)
With the model having predicted a numerical value, additional steps needed to be taken to fully
evaluate the model. These steps included using common evaluation methods called confusion
matrix, classification report and lastly the accuracy score. According to [10] accuracy is a metric
measuring how many cases the model classified correctly divided by the full number of cases;
𝑁𝑢𝑚𝑏𝑒𝑟 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝐶𝑜𝑟𝑟𝑒𝑐𝑡𝑙𝑦
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝐶𝑎𝑠𝑒𝑠

23
Furthermore, [10], states that if all predictions are correct, the accuracy score is 1.0, and 0 if all
cases are predicted incorrectly.While accuracy alone is normally a reliable metric of performance,
it may hide a lopsided number of false-positives or false-negatives. This isn't a problem if there's
a balanced number of false-positives and false-negatives, but this isn't something we can ascertain
using accuracy alone, which leads us to the following two evaluation methods [10]. The two-
evaluation mentioned by Theobald, are confusion matrix and classification report. The two
evaluation methods were imported via:
𝑓𝑟𝑜𝑚 𝑠𝑘𝑙𝑒𝑎𝑟𝑛. 𝑚𝑒𝑡𝑟𝑖𝑐𝑠 𝑖𝑚𝑝𝑜𝑟𝑡𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛_𝑟𝑒𝑝𝑜𝑟𝑡, 𝑐𝑜𝑛𝑓𝑢𝑠𝑖𝑜𝑛_𝑚𝑎𝑡𝑟𝑖𝑥
With the above classes imported the trained model was evaluated by calling the following code:
𝑝𝑟𝑖𝑛𝑡(𝑐𝑜𝑛𝑓𝑢𝑠𝑖𝑜𝑛_𝑚𝑎𝑡𝑟𝑖𝑥(𝑦_𝑡𝑒𝑠𝑡, 𝑚𝑜𝑑𝑒𝑙_𝑝𝑟𝑒𝑑𝑖𝑐𝑡))
Producing the following outcome:
Figure 5. Confusion matrix outcome
The confusion_matrix function from sklearn.metrics is used to compute the confusion matrix,
which is a tabular representation of the predicted labels versus the true labels. It helps in
evaluating the performance of a classification model by showing the counts of true positives, true
negatives, false positives, and false negatives.
𝑝𝑟𝑖𝑛𝑡(𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛_𝑟𝑒𝑝𝑜𝑟𝑡(𝑦_𝑡𝑒𝑠𝑡, 𝑚𝑜𝑑𝑒𝑙_𝑝𝑟𝑒𝑑𝑖𝑐𝑡))
Producing the following outcome:
Figure 6. Classification report outcome where 0 = Benign and 1 = Malicious
While the classification report function is used to generate a text report showing various
classification metrics such as precision, recall, F1-score, and support. It provides a comprehensive
summary of the model's performance for each class in the classification problem.
4.6. Comparative Analysis
The outcome of the confusion matrix reveals that the model correctly classified 911 instances as
negative (TN) and 309 instances as positive (TP). However, it misclassified 37 instances as
negative (FN). In the classification report instance, few notable observations were made, these are
precision, recall, f1-score and support. Precision measures the proportion of correctly predicted
positive instances out of the total instances predicted as positive. For class 0, the precision is 0.96,
indicating that 96% of the instances predicted as class 0 were correctly classified. For class 1, the

24
precision is 1.00, indicating that all instances predicted as class 1 were correctly classified. Recall,
also known as sensitivity or true positive rate, measures the proportion of correctly predicted
positive instances out of the total actual positive instances. For class 0, the recall is 1.00, meaning
that all instances of class 0 in the dataset were correctly identified. For class 1, the recall is 0.89,
indicating that 89% of the instances of class 1 were correctly identified. The F1-score is the
harmonic mean of precision and recall, providing a balanced measure of a model's performance.
For class 0, the F1-score is 0.98, indicating a high balance between precision and recall. For class
1, the F1-score is 0.94, also indicating a good balance between precision and recall. Support refers
to the number of instances in each class. In this case, class 0 has a support of 911 instances, while
class 1 has a support of 346 instances. Accuracy measures the overall correctness of the model's
predictions, considering both true positives and true negatives. In this case, the accuracy is 0.97,
indicating that the model achieved an accuracy of 97% on the entire dataset. The macro average
calculates the average of precision, recall, and F1score across all classes, without considering
class imbalance. In this case, the macro average precision is 0.98, recall is 0.95, and F1-score is
0.96. Lastly, the weighted average calculates the average of precision, recall, and F1-score across
all classes, considering class imbalance. The weight is proportional to the number of instances in
each class. In this case, the weighted average precision, recall, and F1-score are all 0.97.
4.7. Results and Conclusions
Overall Performance:
The model achieves an accuracy of 0.97, indicating that it correctly classifies 97% of the instances
in the dataset. This demonstrates the model's effectiveness in distinguishing between the two
classes (0 = Benign and 1 = Malicious). The high precision and recall values for both classes
indicate a strong ability to correctly identify instances of both positive and negative classes. The
F1-scores for both classes also suggest a good balance between precision and recall.
With the model achieving an accuracy f1-score of 97%, it was then tested on random benign and
malicious queries to evaluate its prediction abilities given the high f1-score.
Figure 7. Results of model being used on malicious query

25
Figure 8. Results of model being used on benign query
5. CONCLUSION
The results from the confusion matrix and classification report indicate that the classification
model performs well in detecting and classifying instances of SQL injection attacks. The model
shows high accuracy, precision, recall, and F1-scores for both the negative (class 0) and positive
(class 1) classes. However, there were a few instances misclassified as negative, leading to false
negatives. These results highlight the model's effectiveness in identifying potential SQL injection
attacks and its ability to contribute to the prevention and security of web applications and
databases.
6. RECOMMENDATIONS
Further analysis and fine-tuning of the model may be required to reduce the occurrence of false
negatives and optimize its performance even further. Additionally, the model's performance
should be validated on larger and more diverse datasets to ensure its robustness and
generalizability in real-world scenarios.
REFERENCES
[1] Lu, D., Fei, J., & Liu, L. (2023). A semantic learning-based SQL injection attack detection
technology. Electronics, 12(6), 1-22. https://doi.org/10.3390/electronics12061344
[2] Yunmar, R. A. (2018). Hybrid intrusion detection system using fuzzy logic inference engine for
SQL injection attack. Kursor, 9(3), 83-93. https://doi.org/10.28961/kursor.v9i3.147
[3] Triloka, J., &Sutedi, H. (2022). Detection of SQL Injection Attack Using Machine Learning Based
on Natural Language Processing. International Journal of Artificial Intelligence Research, 6(2).
[4] Demilie, W. B., &Deriba, F. G. (2022). Detection and prevention of SQLI attacks and developing
compressive framework using machine learning and hybrid techniques. Journal of Big Data, 9(1), 1-
30. https://doi.org/10.1186/s40537-022-00678-0
[5] Daniyal, A., Maha, A., &Suaad, A. (2022, 09). Detection of SQL Injection Attack Using Machine
Learning Techniques: A Systematic Literature Review. Journal of Cybersecurity and Privacy, 2,
764-777.
[6] Vähäkainu, P., &Lehto, M. (2019). Artificial intelligence in the cyber security environment. In
Proceedings of the 14th International Conference on CyberWarfare and Security (pp. 431-440).
Stellenbosch: ICCWS 2019.

26
[7] Satapathy, S., Govardhan, A., Raju, K., & Mandal, J. (2015). SQL Injection Detection and
Correction Using Machine Learning Techniques. Advances in Intelligent Systems and Computing,
435–442.
[8] Halfond, W. G. J., &Orso, A. (2005). AMNESIA: Analysis and monitoring for NEutralizing SQL-
injection attacks. In ASE ‘05: Proceedings of the 20th IEEE/ACM international conference on
automated software engineering (pp. 174-183). https://doi.org/10.1145/1101908.1101935
[9] Zhang, W., Yueqin, L., Xiaofeng, L., Shao, M., Mi, Y., Zhang, H., &Zhi, G. (2022). Deep neural
network-based SQL injection detection method. Security and Communication Networks, 2022, 1-9.
https://doi.org/10.1155/2022/4836289
[10] Theobald, O. (2019). Machine Learning with python. In O. Theobald. Scatterplot Press.
AUTHORS
Valerianus Hashiyana is a Senior Lecturer at Department of Computing, Mathematical &
Statistical Sciences under Faculty of Agriculture, Engineering and Natural Sciences,
University of Namibia. His areas of Research are Cybersecurity, Networking & Security,
AI, IOT, E-Health, Next generation Computing and Educational Technologies.
Tel: +264 812830277, Email: vhashiyana@unam.na / vhashiyana@gmail.com.
Taapopi John Angula is currently a final year masters of information technology student
his research work was based on the development of a structured query language injection
prevention system for the banking sector in Namibia.

Detection of Structured Query Language Injection Attacks Using Machine Learning Techniques

Recommended

More Related Content

Similar to Detection of Structured Query Language Injection Attacks Using Machine Learning Techniques (20)

More from AIRCC Publishing Corporation (20)

Recently uploaded (20)

Detection of Structured Query Language Injection Attacks Using Machine Learning Techniques