Image Features Reordering and Adjustment to Mitigate the Transferability Effect of Adversarial Attacks for Autonomous Driving Cars

In this paper, we introduce a novel adversarial detection framework designed to enhance the security and robustness of DNNs, with a particular focus on applications in autonomous driving systems. Our approach leverages segmentation masks and innovative techniques for image feature reordering and adjustment to detect adversarial attacks effectively. We compre-hensively evaluate the proposed framework under a white-box attack scenario, comparing its performance against the state-of-the-art Feature Squeezing method across eleven adversarial attack techniques and three heterogeneous benchmark datasets. The results demonstrate that our framework achieves detection rates exceeding 96% across all datasets, establishing its efficacy and robustness against diverse attack strategies.