Table 1.
Overview of intrusion detection techniques in CAN.
Fig 1.
Workflow Diagram of the Proposed IDS.
Table 2.
Description of Car Hacking Dataset.
Fig 2.
Proposed DL and Residual Connection based IDS for AV Attack Detection.
Table 3.
Layer-wise specifications of the proposed IDS architecture.
Fig 3.
Confusion matrix of the proposed IDS illustrating classification performance across different attack classes.
Fig 4.
Precision, Recall, F1-Score, and Accuracy Metrics for different attack classes obtained by the proposed IDS.
Fig 5.
Learning curves showing training and validation accuracy (left) and loss (right) of the proposed IDS.
Fig 6.
ROC curves of the proposed IDS for various attack classes evaluated at different decision thresholds.
Table 4.
Training and Testing Accuracy and Loss of the Proposed IDS Across Various Attack Classes.
Table 5.
Performance Metrics of the Proposed IDS on the Validation Dataset Across Various Attack Classes.
Table 6.
Per-seed/per-split class counts and attack rates on the HCRL Car Hacking dataset.
Table 7.
Summary of model performance averaged across three random seeds (11, 42, 99) for each attack type in the HCRL Car-Hacking dataset. Values represent mean ± standard deviation over independently trained models, showing consistent near-perfect detection performance and negligible inter-seed variance.
Fig 7.
Confusion matrices for test splits across multiple seeds and attack types.
Table 8.
Robustness of the proposed IDS under test-time input perturbations (seed = 11). Gaussian noise was added to timestamps, CAN IDs were shifted by ±1, and DATA bytes were jittered by ±1 within [0,255].
Table 9.
Bootstrap 95% confidence intervals (B = 1000) for F1 and ROC–AUC on the held-out test set (seed = 11).
Table 10.
Results of Ablation analysis of the proposed architecture for cyber attack detection in CAV.
Table 11.
Optimized hyperparameters and their corresponding values used in the proposed IDS.
Table 12.
F1-Score comparison of the proposed IDS with existing state-of-the-art methods on the Car Hacking dataset across four major attack types.
Table 13.
Comparison of the proposed IDS with existing techniques in terms of model complexity (number of trainable parameters).
Fig 8.
LIME-based interpretability visualizations for attack instances classified as Class 1 by the proposed IDS.
Fig 9.
LIME visual observations highlighting influential features for attack-free (Class 0) instances detected by the proposed IDS.