Detecting SQL injection attacks by binary gray wolf optimizer and machine learning algorithms

Traditional SQLi methods use blacklists to filter unauthorized characters. There have been numerous works completed in this field. Input validation techniques to avoid SQLi are divided into whitelist validation and blacklist validation techniques [4]. In [5], authors used penetration testing to compensate for the shortcomings of the blacklist filter defense mechanism. Like the blacklisting method, in [6], the authors used the Boyer-Moore string-matching algorithm to calculate how much a string matches the URL with an injected SQL string to determine the existence of a SQL attack. In [7], an effective automated tool is presented to prevent SQLi attacks at run time. In [8], the capabilities of the latest Blackbox web-based security scanners against SQL and XSS injection attacks were investigated. Hsiu-Chuan Huang introduced a new vulnerability scanner for web applications that uses penetration testing to prevent SQLi and detect it. In [9], a method called WAVES was suggested for testing SQLi vulnerabilities in web applications. It uses a web crawler to find points in a web application that can be used for SQLi attacks. The probe then creates attacks that target such locations based on a specific list of attack patterns, using machine learning techniques to improve the method of attack. This solution further improves penetration testing methods by using machine learning approaches to guide testing. However, like all black box testing and penetration testing methods, it cannot provide a guarantee of completeness. A detection mechanism for SQL injection attacks has been proposed in [10] that removes the value from a SQL query attribute of web pages and compares it to a specified value. In this approach, static and dynamic analysis are merged. The experiments demonstrate that the suggested strategy is straightforward and highly successful.

One technique for SQL injection prevention is static analysis of SQL queries in web applications. This method focuses on validating the type of user input to reduce the SQLi probability, rather than detecting them. Carl Gould et al. [11] used the library of Java String Analysis (JSA) to validate the type of user input and prevent SQLi attacks. However, it cannot prevent SQLi attacks if the malicious input data has the correct input type or syntax. Also, the mentioned library only supported the Java language. In [12], the static analysis method was combined with automatic reasoning. This method assumes that there is no tautology in the automatic SQL query generation, and this assumption is also verified. Thus, it was an effective method for detecting SQLi attacks, but except for tautology, it could not detect other SQLi attacks. Stephen Thomas et al. [13] developed a new SQL query by collecting simple SQL statements and executing them to validate the type of user input. The method did not prevent or detect SQLi attacks directly, but it tried to prevent attacks by removing vulnerabilities in the way SQL queries were written. This method was also used for web applications written in the Java language.

The dynamic analysis method examines the web application's response after scanning it. Unlike the static analysis method, it can find the vulnerability of SQL injection attacks without any changes to the web application; so, it has an advantage over the static analysis method. However, the weaknesses found in the web application pages must be manually resolved by the developers. Yuji Kasuga et al. [14] introduced a method called Sania that protects against SQLi attacks. The approach collects natural queries between the client and the web application, as well as between the web application and the database. The authors of [15] proposed a web scan method with advanced features for testing and detecting intrusion into PHP-based web applications; this method can identify the web application's weak point against SQLi attacks. One disadvantage of this tool is that it only works on PHP-based web applications and will not be able to detect SQLi attacks in ASP-based web applications.

This method is a combination of static and dynamic analysis that provides the merits of both methods. Furthermore, it analyzes web pages and creates SQL queries to evaluate the results. William et al. [16] proposed a method called AMNESIA. This method is based on a combination of static and dynamic analysis. In the static stage, it uses the static analysis method to build models of different types of queries that an application can legally generate anywhere it has access to the database. In the dynamic step, it tracks all queries before sending them to the database and checks each query against statistically constructed models. In [17], Buehrer et al. presented a static and dynamic analysis called SQL Guard, which compared static and dynamic SQL queries generated by the user to detect SQLi attacks using a survey tree. SQLCheck is another attack injection based method. Both SQLGuard and SQLCheck methods use a hidden key to determine the user input limit when analyzed by the runtime controller.

SQLrand, which was developed in [18], is a randomization-based method of setting an instruction set that allows developers to use random instructions instead of regular SQL keywords. As a result, a proxy filter inserts queries into the database while removing keywords from normal. Since the attacker's SQL-infused code is not made up of random instruction sets, the injectable commands lead to a syntactically incorrect query. A query-profile-based technique that might identify SQLi attacks without altering the web application was proposed in [18]. However, the web application had to be re-profiled whenever it changed.

Valeur et al. [19] suggested a method based on an intrusion detection system (IDS) using machine learning algorithms. To identify SQLi attacks, this technique employs numerous anomaly identification models. This method as a module is connected to the link between the database server and web applications; the queries made by the application are intercepted and sent to the attack detection system. The intrusion detection system parses the SQL queries and then compares them with the generated model during the execution of the web application to check the differences.

One of the fundamental limitations of learning-based methods is that they cannot always guarantee success in their detection abilities. Because it depends on the quality and method of training used in the presented model, it can cause many false positives and negative results. Joshi et al. [20] presented a method called the Naive Bayes algorithm, which is a classification model in machine learning based on Bayes' theorem. This method assumes that the presence of a feature in a data model is not related to the presence of other features. This method is very simple to implement but could not identify several SQL injection attacks; especially, when a particular type of SQL injection was used for the first time.

In this section of the paper, the required dataset to train the supervised machine learning algorithm is explained. As seen in Fig. 2, the SQL query profiling process is used to create SQLi dataset.

The experiment system consists of some web applications with a frontend, HTTP APIs, and MySQL server backend. The SQLi traffics (normal and SQLi queries) have been generated using SQL query generation application. The resulting MySQL traffic between the web app server and the database server is captured. These collected raw data sets were then processed and correlated to create a separate dataset containing features (labeled records). Table 1 shows the structure of the records in the collected raw dataset. This dataset comprises attacks and normal queries for a web application. These queries were implemented by different SQL structures and complexities. There are simple and complex nested queries in the dataset. Each row in the selected dataset is a normal or malicious SQL query performed in a database of a real-world web application. Each record in the dataset consists of SQL keywords, fragmented text, single quotes, semicolons, comments, intermediate data, and so on.

The raw dataset was filtered, cleaned, and converted to a numeric dataset. This study used a dataset that includes 1027 extracted unique SQL queries (normal and malicious) retrieved from a textual dataset. This dataset includes 1027 records in such a way that each record indicates a SQL query. In this dataset, 554 records are malicious queries, and 473 records are normal queries. The labeling of dataset records is expressed in binary values of 0 (normal SQL query) or 1 (malicious SQL query). Each row of the original dataset consists of two text columns; the first column includes the source of the SQL query, and the second column indicates its label.

Creating an effective numeric training dataset that includes 13 numeric features is the first contribution of this study. The homogeneity of features in a dataset makes machine learning algorithms perform better in terms of accuracy and precision. In the first stage of the proposed methods, the standard SQLi datasets that include thousands SQLi queries were analyzed, and 13 numeric effective features were extracted. In this stage of this study, the selected original dataset was analyzed, and the 13 features were extracted from the source code of each query. All these 13 features are numeric. Twelve features are independent, and the final feature is dependent and indicates the label of the query. The query is normal when its label is zero; the malicious queries are labeled by one. All extracted features have numeric values. Length of query, number of nested queries, number of constants, number of punctuations, and number of logical operators are some of the extracted features. Finally, the created dataset includes 1027 records; each record (each row) includes 13 features extracted from the source code of the queries. The homogeneity of features in a dataset makes machine learning algorithms perform better in terms of accuracy and precision. Table 2 shows the structure of the created dataset. Table 3 describes the features of the created training dataset.

After creating the scalar (numeric) training dataset, in the second stage, two different binary versions of the GWO algorithm were developed. The developed two bGWOs algorithms were used for feature selection. Indeed, the optimal features of the training dataset were selected before creating the desired classification model by the machine learning algorithms. The gray wolf optimization algorithm is a metaheuristic algorithm inspired by the hierarchical structure and social behavior of gray wolves during hunting [21]. It is an algorithm based on population and is simply capable of generalizing to large-scale problems. All the members have a precise hierarchy, and they have certain tasks. In each group of wolves to hunt, there are four levels, which are modeled as a pyramidal structure: the leader wolves are called the alpha (α) group, which can be male or female. These wolves dominate the flock (group). Beta (β) wolves are assisted by alpha wolves in the decision-making process and are also susceptible to being selected instead. Delta (δ) wolves are lower than beta wolves and include old predators. Omega (ω) wolves have the lowest rank in the hierarchy pyramid, which is the least right than the rest of the group; they eat and are not involved in the decision-making process. GWO algorithm consists of three main steps: Tracking and Approaching, encircling, and attacking. In the GWO, Eqs. 1 and 2 were used to mathematically model the encircling behavior of the gray wolves.

In Eqs. 1 and 2, t represents the current repetition, and \(\overrightarrow{A}\) and \(\overrightarrow{C}\) the coefficient vectors, and \(\overrightarrow{X}\) the position vector of a gray wolf, the \(\overrightarrow{A}\) and \(\overrightarrow{C}\) vectors are calculated by Eqs. 3 and 4:

In Eqs. 3 and 4, \(\overrightarrow{a}\) vector components are linearly reduced from 2 to 0 during repetition and \({\overrightarrow{r}}_{1}\) and \({\overrightarrow{r}}_{2}\) vectors are random vectors in the interval \(\left[0, 1\right]\). To see the impact of the relationships mentioned above, a two-dimensional space with possible locations is shown in Fig. 3. A gray wolf with a location \((X,Y)\) can update its location according to the position of the prey \(({X}^{*},{ Y}^{*})\). The locations of the gray wolves were updated based on the (X*, Y*). The Algorithm iteratively converges to the best solution (location of prey). The best solution indicates the minimum number of attributes of the dataset that have maximum effect on the SQLi detection. Consider the vector \(\overrightarrow{A}\) and \(\overrightarrow{C}\) values and the current location, different positions are around the best response. For instance, it is possible to reach a place \(({X}^{*}-X, { Y}^{*})\) according to.

\(\overrightarrow{A}=(\mathrm{1,0})\) and \(\overrightarrow{C}=\left(1, 1\right)\). The two random vectors \({\overrightarrow{r}}_{1}\) and \({\overrightarrow{r}}_{2}\) allow wolves to access all positions. Abstractly, it can be noted that a gray wolf can randomly update its position according to the relationships listed around the prey.

Gray wolves can detect the location of the prey and surround them. The alpha wolves are often the ones that hunt; however, beta and delta wolves occasionally join them. Alpha, beta, and delta wolves are more aware of the probable location of prey; their locations are updated by Eqs. 5, 6, and 7. This allows us to mathematically replicate the hunting behavior of gray wolves. Figure 4 illustrates how a search agent updates its position according to Alpha, Beta, and Delta in the two-dimensional search space. The Pseudo-code of the GWO algorithm is presented in Algorithm 1.

The Continuous Gray Wolves Optimization (OGWO) algorithm constantly shifts their positions to anywhere in the state space. In some special issues, such as selecting features, solutions are limited to binary values of zero and one which creates a binary version [18]. In the method, the wolf Equation is to update a function of three positional vectors, namely, \({\overrightarrow{X}}_{\propto },{\overrightarrow{X}}_{\beta }\, and\, {\overrightarrow{X}}_{\delta }\). It attracts each wolf to the first of the top three solutions. The set of solutions at any given time is binary. To update the positions of a given wolf, following the original gray wolf algorithm, while keeping binary constraints based on Eq. 7, one of the following two approaches can be used:

First approach: In this approach, Eq. 7 can be formulated as Eq. 8. In Eq. 8, \({\text{Crossover}}({\overrightarrow{x}}_{1},{\overrightarrow{x}}_{2},{\overrightarrow{x}}_{3})\) is a suitable shift between x, y, and z solutions. The binary vectors \({\overrightarrow{x}}_{1},{\overrightarrow{x}}_{2} and {\overrightarrow{x}}_{3}\) effectively move wolves toward alpha, beta, and delta wolves. The vectors \({\overrightarrow{x}}_{1},{\overrightarrow{x}}_{2} and {\overrightarrow{x}}_{3}\) are calculated using Eq. 9.

So, the vector \({\overrightarrow{X}}_{1}^{d}\) is the position of the alpha wolf and \({b{\text{step}}}_{\alpha }^{d}\) a binary step in the \(d\) dimension. \({bstep}_{\alpha }^{d}\) is calculated based on Eq. 10. Here, \(rand\) is a random number based on a uniform distribution between zero and one. \({cstep}_{\alpha }^{d}\) is the step size being a continuous value for \(d\) dimension and is calculated using the Sigmoid function (Eq. 11).

\({A}_{1}^{d}\) and \({D}_{\alpha }^{d}\) are calculated using Eqs. 1 and 3. To calculate \({X}_{2}^{d}\) and \({X}_{3}^{d}\) which are the positions of beta and delta wolves, respectively, Eq. 12 is used. So, the \({\overrightarrow{X}}_{2}^{d}\) vector positions are beta wolves and \({bstep}_{\beta }^{d}\) a binary step in the \(d\) dimension. \({bstep}_{\beta }^{d}\) can be calculated based on Eq. 13.

Again, here, \(rand\) is a random number based on a uniform distribution between zero and one. \({c{\text{step}}}_{\beta }^{d}\) is the step size being the continuous value for the \(d\) dimension. It is calculated using the Sigmoid function indicated by Eq. 14. Here, also, \({A}_{2}^{d}\) and \({D}_{\beta }^{d}\) respectively using Eqs. 1 and 3. So that the \({\overrightarrow{X}}_{3}^{d}\) is vector positions of delta wolf (calculated by Eq. 15) and \({bstep}_{\delta }^{d}\) are a binary step in the \(d\) dimension. \({bstep}_{\beta \delta }^{d}\) is calculated based on Eq. 16.

In Eq. 16, a random number \(rand\) is based on a uniform distribution between zero and one and \({cstep}_{\delta }^{d}\) the step size is a continuous value for the \(d\) dimension and is calculated using the Sigmoid function (Eq. 17). In this regard, also, \({A}_{3}^{d}\mathrm{ and }{D}_{\delta }^{d}\) are calculated using Eqs. 1 and 3. \({X}_{d}\) is calculated by Eq. 18. In this relationship, \({a}_{d}\), \({b}_{d}\) and \({c}_{d}\) are binary values for the first, second, and third parameters in dimension \(d\). The \({X}_{d}\) is the output of Crossover in dimension \(d\) and \(rand\) is a random number from a uniform distribution in the range of zero and one. The pseudocode of the bGWO1 is explained in Algorithm 2.

Second Approach: To develop the second version of the bGWO, only the updated gray wolf position vector is forced to be binary. Equation 9 in the first approach is replaced by Eq. 19. So a random number \(rand\) is between zero and one. \({X}_{d}^{t+1}\) is the updated binary position in the dimension \(d\) and in repetition \(t\), as well as the sigmoid function calculated by Eq. 20. The pseudocode of the bGWO2 is explained in Algorithm 3.

In this section, both the first and second approaches of binary GWO algorithms are adapted in feature selection for classification problems. According to the multiplication or permutation principle, for a feature vector of size n, there are \(n{2}^{n}\) different choices of features, which creates a very large and tedious space, and all these \({2}^{n}\) choices must be thoroughly explored. Therefore, the binary gray wolf algorithm is adapted to search the matching feature space to find the best combination of features. It is remembered that the best combination of features is the combination with maximum classification efficiency and minimal selection of the number of features. The Fitness Function, which is used to evaluate the individual positions of gray wolves in the binary gray wolf algorithm, is obtained by Eq. 21.

In Eq. 21, \({\gamma }_{R}\left(D\right)\) is the classification quality of the set of features that is determined relative to the decision\(D\). Also, \(C\) is the total number of features, and \(R\) is the number of selected features, and the ratio \(\frac{\left|C-R\right|}{C}\) is the ratio of unselected features to the total number of features. On the other hand, \(\alpha\) and \(\beta\) are two parameters corresponding to the importance of classification quality and the number of selected features so that\(\alpha \in \left[\mathrm{0,1}\right] , \beta =1-\alpha\). The fitness function maximizes the classification quality\({\gamma }_{R}\left(D\right)\). The ratio of unselected features to the total number of features is indicated by\(\frac{\left|C-R\right|}{C}\). The fitness function is transformed into a minimization function by substituting the error rate for the classification quality and using the number of selected features instead of the number of unselected features (Eq. 22).

In Eq. 22, \({E}_{R}\left(D\right)\) is the error rate for classifying the feature set, \(\left|R\right|\) is the number of selected features, and \(\left|C\right|\) is the total number of features. Also, \(\alpha\) and \(\beta\) are constants to control the classification accuracy and reduce the number of selected features, so that;\(\alpha \in \left[\mathrm{0,1}\right] , \beta =1-\alpha\).

After creating the optimal training dataset that includes optimal features, artificial neural network (ANN) and decision tree (DT) algorithms were used to train the classification model. ANN structures were created from the biological neural systems and the human brain itself. Information processing units, which are called artificial neurons, are the basis of the operation of an ANN. they are the simplified models of brain cells and biological neurons. An artificial neuron can have multiple inputs, but only one output. Artificial neurons used in neural networks are nonlinear and usually provide continuous outputs. Figure 5 shows the mathematical model of a neuron, which is the basis for the design of artificial neural networks.

In Fig. 5, multiple input signals from the external environment are represented by the set \(\left\{{x}_{1}, {x}_{2},{x}_{3},\dots ,{x}_{n}\right\}\). In this study, the input signals are the values of the features in the training dataset. Also, the weighting performed by the synaptic connections of the network is implemented on the artificial neuron as a set of synaptic weights \(\left\{{w}_{1}, {w}_{2},{w}_{3},\dots ,{w}_{n}\right\}\). In the next step, the relevance of each input of the neuron \(\left\{{x}_{i}\right\}\) is calculated by multiplying them by the corresponding synaptic weight \(\left\{{w}_{i}\right\}\). Input signals \(\left({x}_{1}, {x}_{2},{x}_{3},\dots ,{x}_{n}\right)\) are signals (dataset features) that come from the external environment (train dataset). Input signals were optimized by the bGWO to increase the computational efficiency of learning algorithms. Synaptic weights \(\left({w}_{1}, {w}_{2},{w}_{3},\dots ,{w}_{n}\right)\) are used to determine the weight of each feature in the SQLi dataset. Specifically, each feature in the dataset \(\left\{{x}_{i}\right\}\) at the input of synapse j connected to the neuron is multiplied by the synaptic weight \(\left\{{w}_{i}\right\}\). Linear adder (∑) sums all the weighted input features to produce an activation voltage.

Activation Threshold or Bias (b) is a variable used to determine the appropriate threshold. The result produced by the linear adder has a stimulus value toward the output of the neuron. The activation function (φ) is to limit the output of the neuron to a suitable range of values. Typically, the normalized amplitude of the output of a neuron is written with the closed interval [0,1] or with the closed interval [− 1,1]. The output signal (y) contains the final value produced by the neuron by a particular set of input features that can be used as input to other connected neurons. Equation 23 was used to calculate the amount of output signal. In this regard, the activation function is such that the function is \(\mathrm{\varphi }\). \(\varphi :{\mathbb{R}} \to {\mathbb{R}}\) in the space of real numbers and explains the neuronal output. The activation function in an artificial neuron acts so that the output of the neuron (y) in a neural network is between special values (usually between 0 and 1, or between − 1 and 1). The activation function can be bounded by Eq. 24.

A decision tree is a hybrid data structure of the graph and tree structures; Where the internal node represents the preparation of the feature, and each branch represents the test result and each leaf node represents the class label. Also, the paths from the head to the leaves show the classification rules. Typically for decision analysis, decision trees are regularly used in data mining to help identify a strategy that is most likely to achieve a goal.

To evaluate the proposed method, firstly, the required scalar (numeric) training dataset was created. Then, the proposed feature selection bGWO algorithms were implemented in MATLAB. The ANN and DT machine learning algorithms were implemented in two forms; In the first form, the machine learning algorithms were implemented without using selective features. In the second form, the machine learning algorithms were implemented using feature selection algorithms. Indeed, in the experiments, two different classification models were created. In the first form of implementation, the ANN and DT train the classification model using the dataset with all features. In the second implementation, the machine learning algorithms invoke the bGWO for selecting the optimal features; then, the machine learning algorithms train the classification model using the dataset with optimal features. The hardware and software specifications of the implementation environment are listed in Table 4.

In this study, the SQLi training dataset has been used to train and test the ANN and DT algorithms.

In this study, 70% of the dataset was used for training the classification model and 30% of the dataset was used for testing the created classification model. In the test stage, a subset (30%) of the training dataset was also used. The selected training and test datasets have uniform distribution and consist of normal and malicious queries. Figure 6 shows the raw dataset that includes the source code of the SQL queries (normal and malicious).

These features are defined based on the important features in SQL injection and a weight is assigned to each feature based on their importance. In this research, nominal features such as the length (number of words) of the query, the number of UNION commands, the number of special characters, the number of AND operators, the number of parentheses, and other features have been used. To improve the performance of the classification models, it is more desirable to use numerical features in the training dataset. Hence, the extracted nominal features from the raw dataset were converted into numerical features. Figure 7 shows the numeric form of the training dataset. As shown in Fig. 7, features are selected for each query, and each row of the dataset indicates the numerical features of a query. Also, the last column of this table indicates that each row (query) is a normal or malicious query. In this research, the converted dataset has been used to train and test the machine learning algorithms. Table 5 shows the specifications of the numeric training dataset.

To implement the proposed method, a program was implemented in the MATLAB programming environment version 2020b. Accuracy, precision, and sensitivity are the classification criteria that were used to evaluate the proposed method; these criteria were applied as a fitness function of the binary gray wolf algorithm. The parameters of the developed binary gray wolf algorithm have been calibrated experimentally during the experiments. Table 6 shows the best values of the bGWO parameters.

The proposed algorithm is implemented in two scenarios named bGWO1 and bGWO2. The scenarios are performed on a dataset to determine the efficiency criteria for the neural network and the decision tree algorithms. The performance criteria that were used in this study are as follows:

These criteria were applied as a fitness function of the binary gray wolf algorithm. Extensive experiments were conducted on the created data set that answers the research questions as follows:

In this study, two series of ML experiments have been performed; in the first experiment, the ANN and DT were applied to the whole dataset (13 features). In the second experiment, the ANN and DT were applied in the filtered dataset (with the selective features by the proposed bGWO). The results of the experiments have been analyzed to evaluate the effectiveness of the proposed feature selector in the SQLi detectors’ performance.