This paper investigates the interpretability of various machine learning models in the context of detecting SQL injection attacks. The models under investigation include decision trees, multi-layer perceptron, random forests, support vector machines, ADA Boost, naive Bayes, and quadratic discriminant analysis. Our main objective is to gain insights into how these models make decisions when classifying SQL commands as either legitimate or potentially malicious. To achieve this, we employ a set of commonly used features from the literature to train and evaluate the models. By prioritizing explainability, our aim is to uncover the underlying factors and decision rules that drive the models' predictions. Through this research, we seek to bridge the gap between model accuracy and human understanding, thereby facilitating the practical application of machine learning models in real-world security scenarios. The ifndings of this study contribute to the development of a case-based explanation system that provides valuable insights for cybersecurity practitioners and researchers.
The increasing reliance on web applications and the growing sophistication of cyber threats
have necessitated the development of efective methods for detecting and mitigating SQL attacks.
Machine learning (ML) models have shown promising results in identifying such attacks by
leveraging their ability to capture complex patterns and relationships in data [
Explainability is a critical aspect in the context of cybersecurity [
Understanding the decision-making process of ML models is crucial for building trust and
confidence in their outputs [
In this paper, we aim to investigate the explainability of several ML models for the detection
of SQL injection attacks, particularly, DTs, multi-layer perceptron (MLP), random forests (RFs),
support vector machines (SVMs), ADA Boost, naive Bayes (NB) and quadratic discriminant
analysis (QDA). Our objective is to shed light on the interpretability of these models when
attempting to classify a dataset of SQL commands, both legitimate and considered as possible
attacks, by using a set of features commonly found in the literature [
This paper is organized as follows: Section 2 provides an overview of related work on explainable ML models and the detection of SQL attacks. Section 3 describes the dataset and evaluates the performance of several state-of-art ML models for the classification of SQL attacks. We then present in Section 4 our methodology for the explainability of these models, discussing the insights gained from the application of the explainable artificial intelligence (XAI) techniques. Subsequently, in Section 5 we present a case-based explanation approach that exploits the conclusions of the XAI analysis to provide useful explanation examples to the end user. Section 6 analyzes and evaluates its performance. Concluding remarks are discussed in Section 7, including a summary of our findings and implications.
Cybersecurity is in charge of protecting everything that can connect to a diferent digital device,
it can be divided into two parts, the systems’ security information and the physical security of
any devices [
One of the most common types of cybersecurity attacks are injection attacks. Injection attacks
happen when an attacker makes an input which the system is not ready to receive; this allows
the attacker to execute malicious commands which can result in a diversity of negative efects
such as data leakage or giving access to users who should not be authorized to have them [
However, it is considered that the more powerful a ML model is, the less interpretable
and explainable it becomes. This issue is particularly important in the field of cybersecurity
considering that fundamental decisions cannot be trusted to a system that cannot explain
itself. This led to the field of XAI, which aims to shed light on intelligent systems without
compromising their performance. Explanability attempts to solve the question as to why a
model reaches a given conclusion, since some models are used in a high-risk enviroment, it
is important to comprehend and not just blindy trust it because it is possible the model might
not have enough information to understand the situation as a whole and give an incorrect
prediction [
XAI techniques in cybersecurity have focused on the following fields: Intrusion Detection
Systems, Malware Detection, Phishing and Spam Detection, and BotNet Detection [
However, other cybersecurity fields have extensibly been favored by XAI techniques. For
instance, the explainability of Intrusion Detection Systems has been explored using rule-based
models [
The dataset of SQL Injection Attacks was obtained from [
DT Attack).
From the SQL sentence 82 binary and numerical features were obtained through the analysis
of the existing literature [
Next, diferent ML classification models were trained and evaluated using 5-fold crossvalidation. Similarly to other previous results from the literature, they all achieved very high accuracy, as presented in Table 1. These results led us to initially presume that there was a direct correlation between some of the features and the target label. Thus, the correlation table was computed between all the features (including the target variable). This table is presented in Figure 1 where column/row 0 represents the target variable. As we can observe, there is no correlation between any of the dataset’s features. Another possible explanation for this remarkable accuracy is that the combination of values from a bounded subset of features leads to a clear classification. To discard this hypothesis, we analyzed the decision tree model, as is a white-box ML technique. The corresponding tree is presented in Figure 2, which clearly led us to reject this hypothesis.
After discarding any correlation or simple direct proportion between features that led to the classification as an attack, we can conclude that the ML models are very suitable tools for this task, as they achieve remarkable accuracy even though this is not a simple classification problem. In the following section, we introspect these models with XAI techniques to infer which are the most relevant features for detecting SQL injection attacks.
Explainable Artificial Intelligence (XAI) plays a crucial role in providing transparency to blackbox ML models. While often robust and accurate, these models can be challenging to understand due to their complex internal mechanisms. This is the case of the ML models trained in the previous section: they achieve very high accuracy, although there is no information about which features are the most relevant to identify SQL injection attacks.
XAI techniques aim to bridge this gap by shedding light on how these models make predictions or decisions. By using methods such as feature importance analysis, rule extraction, or local explanations, XAI allows users to gain insights into the factors that influence the model’s output. This transparency not only helps in understanding the decision-making process but also assists in identifying biases, errors, or potential vulnerabilities.
To provide transparency to the previous ML models, we have focused on the following established XAI methods: feature importance, permutation importance and Shapley Additive exPlanations (SHAP). As most of the ML techniques evaluated in the previous section achieve a similar performance (with a diference of 0.5%) we have chosen the Random Forest model for its analysis. The decision tree structure of Random Forests provides inherent interpretability, as the paths from the root to the leaves can be traced to understand how the model arrives at its predictions.
The feature importance method computes the relative contribution of diferent input features on the predictions made by a machine learning model. It aims to provide insights into which features are most influential in driving the model’s output. The feature importance in Random Forests can be easily obtained by analyzing the average impurity reduction (Mean Decrease in Impurity, MDI) across all trees. The permutation importance method evaluates the feature’s impact on the model’s performance when their values are randomly shufled. The logic behind this method is that significant features, when shufled, will cause a more significant drop in the model’s performance, indicating their crucial role in making accurate predictions. Extending this concept, the SHAP (Shapley Additive exPlanations) method is based on cooperative game theory and considers the interactions and dependencies between feature subsets to provide a fair allocation of the model’s prediction to individual features.
Applying these explanation methods to the RF model, we obtain the results displayed in Figure 3, where we have selected the 15 most relevant features for each method. As we can observe, F8 (’from’ keyword) is the most relevant feature to classify a SQL sentence as an attack, followed by F16 (=) and F81 (hidden equality expression). From the intersection of the resulting feature lists, we obtain the 11 most significant explanation features: ℱ = { 1, 2, 6, 8, 16, 24, 25, 52, 57, 81, 82}1 from the whole original set ℱ with 74 features used to train the model.
The next step is analyzing the impact (positive or negative) of the values of these most significant features. To do so, we compute the class distribution ( Attack or No Attack) of the instances having a concrete feature value. As all features in ℱ are boolean values representing 1see Table 2 for the corresponding description that the feature is found or not found in the SQL sentence, we obtain the two charts presented in Figure 4. These charts represent the percentage of instances (among the total) considered as an attack or not when a feature is found (left chart) or not (right chart). Through the class distributions reported in these charts, we can understand the most significant patterns or rules found by the random forest classifier. For example, the presence of a hidden equality expression (F81) such as “1 = 1”, that is commonly used to masquerade an attack, has a majority percentage of instances labeled as attacks (Figure 4 left). Another usual keyword in SQL attacks is sleep (F57), that, when found, also presents a majority percentage of sentences considered an attack. On the other hand, the absence of common keywords such as from (F8) or select (F1) is also a clear indication of a potential attack (Figure 4 right).
After the analysis of the most relevant features of the dataset for the classification of SQL attacks by an ML model, we can move forward to developing a case-based explanation system. By drawing parallels between the current query and similar explanation cases from the dataset, case-based explanations provide a transparent rationale for the ML classification, enhancing the interpretability of the underlying model. In our concrete scenario, SQL sentences similar to a query are presented to the user. These explanation cases corroborate the prediction made by the ML model regarding the malicious nature of the query.
We propose a case-based explanation process following the MAC/FAC (Many-Are-Called,
Few-Are-Chosen) schema [
() = { ∈ : . = () ∧ . = ., ∀ ∈ ℱ}
The selection step (FAC) obtains the most similar explanation cases to using a similarity metric that compares the remaining features in the description of a SQL sentence ℱ . This step applies a minimum similarity threshold and the inverse of the Euclidean distance as the similarity metric:
() = (, ) = { ∈ () : (, ) > } 1 − √︁∑︁(. − . )2, ∀ ∈ ℱ (1) (2) (3)
In the following section, we present the evaluation of the proposed case-based explanation method following the MAC/FAC approach.
To evaluate the case-based explanation method we have split the dataset into a 25%-75% test and training sets. The filtering step (MAC) is evaluated by computing how many explanation cases from the training set are obtained, this is, how many match with the query for the most significant explanation features ℱ. This process is repeated for each case in the test set. Results are shown by the histogram in Figure 5 (left). As we can observe, the histogram shows that the MAC filter is able to find a large number of explanation cases obeying the restrictions given by the explanation features in ℱ. This is mainly due to the very large number of instances in the dataset To evaluate the selection step (FAC) we computed the mean similarity of the k-NNs to the query. Figure 5 (right) shows the corresponding histogram for = 3. As we can observe, our case-based explanation system is able to find explanation cases that are very similar to the query. Again, this is rooted in the large number of options given by the previous MAC step.
This paper reports an in-depth analysis of the performance of ML models for detecting SQL injection attacks. The first contribution of the paper is the analysis and collection of the features considered by previous works for training such models. Then a novel dataset represented by these features is developed from a large collection of SQL sentences labeled according to their malicious intents. From this dataset, several ML models are trained and evaluated, demonstrating the high accuracy of these techniques for identifying SQL attacks. However, the black-box nature of these ML models does not allow us to understand the most significant features that denote a SQL attack. Therefore, we apply several XAI techniques to identify the features that characterize malicious SQL sentences. Finally, these relevant features let us build a case-based explanation system based on the MAC/FAC schema. This CBR system enables the understanding of why a SQL sentence is considered an attack by the ML model through the comparison between such sentences and similar explanation cases. count(SUBSTRING) count(MID) count(ASC) count(<) count(>) count(") count("") exists floor rand group order lenght ascii if count sleep between values delay wait benchmark indentity_insert truncate table username | password user | pass ")" limit concat ne find eq | gt | gte | lt | it | lte | ite | in | nin mod | regex | text return return true | return 1 exists create show collection while(loop) hidden equality expression large expression Feature
Supported by the PERXAI project PID2020-114596RB-C21, funded by the Ministry of Science and Innovation of Spain (MCIN/AEI/10.13039/501100011033) and the BOSCH-UCM Honorary Chair on Artificial Intelligence applied to Internet of Things.