Adversarial Attack Detection via Multi-Task Consistency in Vision Systems

A new research paper on arXiv (2608.07750) proposes an efficient method for detecting adversarial attacks in deep neural networks (DNNs) used in vision perception systems, particularly for autonomous driving. The method leverages multi-task perception, where inconsistencies between outputs of different vision tasks (e.g., object detection and instance segmentation) signal the presence of adversarial perturbations. The authors developed a consistency score metric to measure these inconsistencies and designed an approach to select optimal model pairs for effective detection. The approach aims to be cost-efficient, addressing the impracticality of existing defenses for resource-constrained applications. The paper was announced as a cross-type submission on arXiv. The research addresses growing concerns about the vulnerability of DNNs in safety-critical applications like autonomous driving, where adversarial attacks could lead to catastrophic failures. The proposed detection scheme offers a promising direction for enhancing the robustness of such systems without significant computational overhead.