This section includes the evaluation measurements used to evaluate the model performance and explains the findings and results of the proposed model in detail. The efficiency of this study’s proposed model was evaluated utilizing different evaluation measurements, including accuracy, precision, recall, and F1-score and confusion matrix. Evaluation measurements were utilized to assess the prediction and classification problems.
4.1. Machine Learning Approach
Table 2 presents the performance results for the classification report for an RF classifier. Adversarial Benign had a precision of 0.61, which means 61% of instances predicted were correctly identified, and the recall value was 0.63, which means the classifier identified 63% of all actual instances of this class in the dataset. The F1-score, which balances precision and recall, for this class was 0.62. Adversarial Noise Injection had a precision of 0.73, which means 73% of instances were predicted correctly. The recall value was 0.68, which means the classifier identified 68% of all actual instances, and the F1-Score for this class was 0.70. Adversarial Outlier had a precision of 0.66, which means 66% of instances predicted as Adversarial Outlier were correct. The recall was 0.68 for this class, the classifier identified 68% of all actual instances, and the F1-score was 0.67. Adversarial Perturbation had a precision of 0.82. A total of 82% of instances predicted were correct, with a recall of 0.83 for all actual instances, and the F1-score for this class was 0.83. The Adversarial Perturbation class had the highest precision, recall, and F1 score. However, the performance for the Adversarial Benign class was the lowest among the four classes, with the lowest precision, recall, and F1 score. The model’s overall accuracy was 0.70, meaning that it correctly classified 70% of all instances across all classes. The macro-average values provide an average across all classes, with precision, recall, and F1-score being 0.71, 0.70, and 0.70, respectively. The weighted average values consider the class distribution, providing a more representative overall performance measure. In this case, the weighted average precision was 0.71, and the recall and F1-score had the same value of 0.70, reflecting the model’s performance while considering the varying numbers of instances in each class. In a classification report, “support” refers to the number of true instances for each label in the dataset. It indicates the number of occurrences of each class in the dataset that was used for evaluating the classifier.
Table 3 presents the performance results for the classification report for the XGB classifier. Adversarial Benign had a precision, recall and F1-Score of 0.55, which means this classifier correctly identified 55% of instances predicted. Adversarial Noise Injection had a precision of 0.70, which means 70% of instances were predicted correctly. The recall value was 0.62, which means the classifier identified 62% of all actual instances, and the F1-Score for this class was 0.66. Adversarial Outlier had a precision of 0.62, which means 62% of instances predicted as Adversarial Outlier were correct. Recall was 0.71 for this class, the classifier identified 71% of all actual instances, and the F1-score was 0.66. Adversarial Perturbation had a precision of 0.82. A total 82% of the instances predicted were correct, with a recall of 0.78 for all actual instances, and the F1-score for this class was 0.80. The Adversarial Perturbation class had the highest precision, recall, and F1 score, and the Adversarial Benign class was the lowest among the four classes, with the lowest precision, recall, and F1 score. The classifier achieved an accuracy of 0.66, suggesting reasonable overall performance in classifying instances correctly. The macro-average values provide an average across all classes, with precision, recall, and F1-score being 0.67, 0.66 and 0.67, respectively. The weighted average values consider the class distribution, providing a more representative overall performance measure. In this case, the weighted average precision was 0.67, and the recall and F1-score had the same value of 0.66, reflecting the model’s performance while considering the varying numbers of instances in each class.
Table 4 presents the performance results for the classification report for the KNN classifier. The Adversarial Benign class achieves a precision of 0.50, a recall of 0.62, and an F1-score of 0.55. This suggests that the classifier identified a relatively high proportion of actual instances. For the Adversarial Noise Injection class, a precision of 0.66, a recall of 0.60, and an F1-score of 0.63 are achieved. This indicates a moderate performance in identifying instances of this class with a balance between precision and recall. The KNN classifier achieves a precision of 0.67, a recall of 0.59, and an F1-score of 0.63 for the Adversarial Outlier class. Similar to the Adversarial Noise Injection class, there is a moderate balance between precision and recall. The Adversarial Perturbation class achieves a relatively higher performance with a precision of 0.80, recall of 0.76, and an F1-score of 0.78. This indicates that the classifier is quite effective in identifying Adversarial Perturbation. The Adversarial Benign class is the lowest among the four classes, with the lowest precision, recall, and F1 score. The overall accuracy of the K-Neighbors classifier is 0.64, suggesting that 64% of all instances are classified correctly across all defined classes. The macro-average precision, recall, and F1-score are 0.66, 0.64, and 0.65, respectively. The weighted average precision, recall, and F1-score are also 0.66, 0.64, and 0.65, respectively.
Table 5 presents the performance results for the classification report for the DT classifier. The Adversarial has a precision of 0.60, meaning that out of all instances, 60% are predicted correctly. The recall and F1-score of 0.59 indicates that the model correctly identifies 59% of all actual instances. For the Adversarial Noise Injection class, the model performs better with a value of 0.68 for all evaluation metrics, including precision, recall and F1-score. The classifier achieves a precision of 0.64, a recall of 0.63, and an F1-score of 0.63 for the Adversarial Outlier class. In the Adversarial Perturbation class, the classifier demonstrates relatively high performance with a precision of 0.78, a recall of 0.81, and an F1-score of 0.80. This indicates that the classifier is quite effective in identifying instances of Adversarial Perturbation. The model’s overall accuracy is 0.68, meaning that it correctly classifies 68% of all instances across all classes. The macro-average values provide an average across all classes, with a 0.68 precision, recall, and F1-score value. The weighted average precision, recall, and F1-score are also 0.67, 0.68, and 0.68, respectively. A weighted average considers the support for each class, giving more weight to classes with a larger number of instances.
Figure 6 visualizes the confusion matrix (CM) of the ML approaches. This visualization offers a high-level overview of how the classification algorithm executes. The performance of this approach appears to be superior, as indicated by fewer false positive and false negative outcomes, along with higher counts of true positive and true negative values. The CM illustrates instances where the algorithm misclassified records, while the diagonal elements represent correct predictions. Overall, the ML approach demonstrates improved accuracy and effectiveness in classification tasks compared to traditional methods.
4.2. Machine Learning Ensemble Approach
Table 6 presents the performance results for the classification report for the ensemble voting classifier. Adversarial Benign has a precision of 0.59, indicating that out of all instances predicted as Adversarial Benign by the classifier, only 59% are correct. It has a recall of 0.64, and the classifier identifies 64% of all instances of adversarial benign behavior in the dataset. F1-score is 0.62, which is the harmonic mean of precision and recall. Adversarial Noise Injection has a precision of 0.74, which means that 67% of instances predicted as Adversarial Noise Injection are correct, and recall with 0.67 and F1 score for this class is 0.70. Adversarial Outlier has a precision of 0.66, in which 66% of instances predicted as Adversarial Outlier are correct. The recall of this class is 0.69, in which the classifier identifies 69% of all actual instances of Adversarial Outliers in the dataset. The F1-score, which balances precision and recall for this class, is 0.68. Adversarial Perturbation has a precision of 0.84, which indicates that 84% of instances are predicted correctly, and recall has a 0.82 value, which the classifier identifies as 82% of all actual instances. The F1 score for this class is 0.83. The Adversarial Perturbation class has the highest precision, recall, and F1 score. However, the performance for the Adversarial Benign class is the lowest among the four classes, with the lowest precision, recall, and F1 score. The model’s overall accuracy is 0.70, meaning that it correctly classifies 70% of all instances across all classes. The macro-average values provide an average across all classes, with a precision and F1-score of 0.71 and recall of 0.70, respectively. The weighted average values consider the class distribution, providing a more representative overall performance measure. In this case, the weighted average precision is 0.71, and recall and F1-score are 0.70, reflecting the model’s performance while considering the varying numbers of instances in each class.
Figure 7 visualizes the CM of the ensemble voting approach. This visualization offers a high-level overview of how the classification algorithm executes. The performance of this approach appears to be superior, as indicated by fewer false positive and false negative outcomes, along with higher counts of true positive and true negative values. The CM illustrates instances where the algorithm misclassified records, while the diagonal elements represent correct predictions. Overall, the ensemble approach demonstrates improved accuracy and effectiveness in classification tasks compared to traditional methods.
4.3. Deep Learning Approach
Table 7 presents the performance results for the classification report for an MLP classifier. Adversarial Benign had a precision of 0.40, indicating that out of all instances predicted as Adversarial Benign by the classifier, only 40% were correct. It had a recall of 0.46, and the classifier identified 46% of all instances of adverse benign behavior in the dataset. F1-score was 0.43, which is the harmonic mean of precision and recall. For this class, it was 0.43, which balances both precision and recall. Adversarial Noise Injection had a precision of 0.67, which means that 67% of instances predicted as Adversarial Noise Injection were correct, and recall with 0.52 indicates the classifier identified 52% of all actual instances of Adversarial Noise Injection in the dataset. The F1 score for this class was 0.58. Adversarial Outlier had a precision of 0.58, in which 58% of instances predicted as Adversarial Outlier were correct. The recall of this class was 0.66, in which the classifier identified 66% of all actual instances of Adversarial Outliers in the dataset. The F1-score, which balances precision and recall for this class, was 0.62. Adversarial Perturbation had a precision of 0.79, which indicates that 79% of instances were predicted correctly, and recall had a 0.73 value, which the classifier identified as 73% of all actual instances. The F1 score for this class was 0.76. The Adversarial Perturbation class had the highest precision, recall, and F1 score. However, the performance for the Adversarial Benign class was the lowest among the four classes, with the lowest precision, recall, and F1 score. The overall accuracy of the model was 0.59, meaning that it correctly classified 59% of all instances across all classes. The macro-average values provided an average across all classes, with precision, recall, and F1-score being 0.61, 0.59, and 0.60, respectively. The weighted average values consider the class distribution, providing a more representative overall performance measure. In this case, the weighted average precision, recall, and F1-score were all 0.61, reflecting the model’s performance while considering the varying numbers of instances in each class. These metrics give equal weight to each class, regardless of its distribution in the dataset.
Table 8 presents the performance results for the classification report for the DNN classifier. The DNN classifier achieves a precision of 0.39 and a recall of 0.44 for the Adversarial Benign class, resulting in an F1 score of 0.41. This indicates that the classifier correctly identifies 44% of instances belonging to the Adversarial Benign class out of all instances. For the Adversarial Noise Injection class, the classifier demonstrates a precision of 0.73, a recall of 0.49, and an F1-score of 0.58. This suggests that while the classifier achieves high precision, it may miss a significant portion of actual instances of Adversarial Noise Injection. The classifiers achieve a precision of 0.57, a recall of 0.70, and an F1-score of 0.63 for the Adversarial Outlier class. For the Adversarial Perturbation class, the classifiers demonstrate relatively high performance with a precision of 0.78, a recall of 0.74, and an F1-score of 0.76. This indicates that the classifier effectively identifies Adversarial Perturbation. The overall accuracy of the DNN classifier is 0.59, suggesting that 59% of all instances are classified correctly. The macro-average precision, recall and F1-score are 0.62, 0.59, and 0.60, respectively. The weighted average precision, recall, and F1-score are also 0.61, 0.59, and 0.59, respectively.
Figure 8 visualizes the CM of the DL classifiers. This visualization offers a high-level overview of how the classification algorithm executes. The performance of this approach appears to be superior, as indicated by fewer false positive and false negative outcomes, along with higher counts of true positive and true negative values. The CM illustrates instances where the algorithm misclassified records, while the diagonal elements represent correct predictions. Overall, the MLP and DNN classifiers demonstrate improved accuracy and effectiveness in classification tasks compared to traditional methods.
4.4. TAAD-CAV Results
Table 9 presents the performance results for the classification report for the meta classifier. The meta classifier achieves a precision of 0.59 and a recall of 0.64 for the Adversarial Benign class, resulting in an F1-score of 0.62. This indicates that 59% of instances predicted as Adversarial Benign are correct, and the classifier identifies 64% of all actual instances of Adversarial Benign in the dataset. For the Adversarial Noise Injection class, the meta classifier demonstrates a precision of 0.74, a recall of 0.67, and an F1-score of 0.70. This indicates that the classifier performs relatively well in identifying Adversarial Noise Injection instances. The classifier achieves a precision of 0.67, a recall of 0.69, and an F1-score of 0.68 for the Adversarial Outlier class. For the Perturbation Injection class, the classifier demonstrates a relatively high performance with a precision of 0.84, recall of 0.82, and an F1-score of 0.83. This indicates that the classifier is quite effective in identifying instances of Perturbation Injection. The macro-average precision, recall, and F1-score are 0.71, 0.70, and 0.71, respectively. This indicates the average performance across all classes without considering class imbalance. The weighted average precision, recall, and F1-score are also 0.71, 0.70, and 0.71, respectively. The weighted average considers the support for each class, giving more weight to classes with a larger number of instances. The meta classifier demonstrates relatively high performance across all classes, with balanced precision and recall. The classifier achieves an overall accuracy of 70%, indicating its effectiveness in classifying instances correctly across all defined classes.
Figure 9 visualizes the CM of the hybrid ML and DL approach, which we named the Meta model. This visualization offers a high-level overview of how the classification algorithm executes. The performance of this approach appears to be superior, as indicated by fewer false positive and false negative outcomes, along with higher counts of true positive and true negative values. Overall, the Meta model demonstrates improved accuracy and effectiveness in classification tasks compared to traditional methods.