Resilience Analysis of Electromagnetic Side-channels against Weight Perturbations

In this paper, we explore the use of electromagnetic (EM) side-channel signal analysis to detect attacks on neural networks (NNs) operating on GPUs, specifically focusing on fault injections and NN parameter perturbation. Our primary discovery reveals that perturbing the weights using different distributions has distinct effects on various neural network architectures. For instance, randomly zeroing some weights impacts the emanated EM signals locally, while using a Gaussian distribution for perturbation affects the signals more broadly. Additionally, we find that employing simple, state-of-the-art signal processing techniques is sufficient for detecting these perturbations, suggesting that EM-based detection systems can reduce latency and enhance accuracy. To facilitate this study, we initially train a neural network using transfer learning methods, then apply different perturbation techniques to various neural layers, and finally, evaluate the detection performance with a straightforward, state-of-the-art signal processing algorithm. This version maintains the integrity of the technical details while enhancing readability and flow, making it more accessible for the intended audience.