RAID: Real-Time XAI-Guided Adversarial Attacks for Edge-Deployed Neural Networks

Interpretability in deep learning is essential for understanding and improving the reliability of model behavior, particularly on edge devices deployed in safetycritical environments. Grad-CAM highlights class-relevant regions in input images and is widely used for model interpretability. However, its computational overhead limits real-time use. We address this by training a lightweight network to approximate Grad-CAM masks, which are then exploited to launch a novel, targeted sticker-pasting attack that manipulates model attention. To enable real-time adversarial testing, we train a Fully Convolutional Network (FCN-ResNet50) to segment and predict Grad-CAM hotspot regions on the fly, allowing seamless attack deployment in real-time. Our pipeline is deployed on NVIDIA Jetson Orin, where it achieves efficient segmentation and real-time attack generation with an average latency of 0.09s per attack. Experimental results across multiple stateof-the-art classifiers, including VGG16, ResNet50, ViT-B16, and EfficientNet, demonstrate significant drops in Top-1 accuracy (e.g., from 96.5% to 16.0% on ResNet50) under our targeted attacks, outperforming random sticker and patch attacks in effectiveness. Furthermore, our attack framework shows strong transferability across diverse model architectures, highlighting a broader vulnerability in saliency-guided interpretation. These findings raise concerns about the malicious use of XAI tools and emphasize the need for robust defense mechanisms in real-time, resource-constrained environments.