With deep learning-based object detectors widely deployed as visual components in Industrial Internet of Things (IIoT) devices like cameras, their adversarial robustness has become paramount to the security and resilience of hyperconnected industrial systems. Existing adversarial defenses are often inadequate for the complexities of object detection, and securing already deployed detectors with a lightweight defense that avoids costly retraining remains a major challenge. In this paper, we propose XAIAD-YOLO: Explainable AI-Guided Adversarial Defense for YOLO detectors, a novel test-time defense to enable resilient YOLO detectors. XAIAD-YOLO introduces a synergistic two-stage purification framework grounded in distinct theoretical principles. Its initial stage, based on signal processing principles, filters high-frequency adversarial noise from genuine image structures. The second stage performs targeted feature destabilization; guided by our efficient XAI saliency map and grounded in the principle of differential feature stability, it precisely neutralizes fragile adversarial artifacts. Experiments show that our XAI method achieves 66.08 FPS (1.56x faster than Grad-CAM++), and our defense method significantly improves adversarial robustness, making anchor-based, anchor-free, lightweight, and non-lightweight YOLO detectors more resilient in both white-box and black-box scenarios. By uniquely integrating explainability into the defense mechanism, XAIAD-YOLO provides a practical and effective solution for enhancing the resilience and trustworthiness of AI in critical industrial applications. Our source code and datasets are available https://anonymous.4open.science/r/XAIAD-YOLO-B0A3/here.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号