Poster Abstract: Mobile Vision Dynamic Layer Dropping against Adversarial Attacks

Deep neural networks (DNNs) have achieved notable success in mobile vision tasks, yet they show vulnerability to adversarial attacks. When carefully crafted perturbations are introduced, these models can be easily misled into wrong classifications, posing significant risks for safety-critical mobile systems like autonomous vehicles. Although various defense strategies, both static and dynamic, have been proposed, many fail to address adaptive attacks or overlook the resource constraints of mobile systems. To address these limitations, in this paper, we present GuSoDrop, a lightweight dynamic defense framework that applies stochastic layer dropping. GuSoDrop leverages randomness to counteract adaptive attacks while selectively dropping less important layers to reduce computation overhead. Our preliminary evaluation shows that GuSoDrop outperforms state-of-the-art defense methods against different adaptive attacks and improves efficiency in reducing computational overhead.