A Comparative Analysis of CNN and Vision Transformer Object Detectors under Physical Adversarial Attacks in Automated Driving

This study examines the robustness of two object detection architectures, a convolutional neural network (YOLO11) and a vision transformer (RT-DETR), under adversarial patch conditions. Both models were evaluated on the APRICOT dataset with patch perturbations using standard detection metrics and custom robustness measures. YOLO11m achieved an [email protected] of 0.15 and [email protected] of 0.12, with 123 misclassifications and 1,052 robust detections. RT-DETR-50 obtained slightly higher accuracy ([email protected] = 0.17, [email protected] = 0.13) and precision (0.317), while producing 406 misclassifications and 1,417 robust detections. These results reveal a trade-off between the two approaches: the vision transformer provides stronger overall detection performance and improved robustness in preserving object recognition despite adversarial interference, whereas the convolutional neural network shows greater resilience to misclassification. The findings emphasise the importance of evaluating both accuracy and robustness metrics when comparing detection architectures for deployment in safety-critical domains.