An Input Recovery Side-Channel Attack on Dnn Accelerator with Three-Dimensional Power Surface

Amid the rapid evolution of diverse Deep Neural Network (DNN) accelerators, the potential side-channel leakage poses a serious threat to the security of Artificial intelligence (AI) devices. This hardware vulnerability will leak sensitive input data of DNN models, allowing adversaries to steal individual privacy and trade secrets. We propose a novel Side-Channel Attack (SCA) methodology leveraging the Three-Dimensional Power Surface (3DPS) for DNN accelerators. 3DPS facilitates structured feature reconstruction of matrix multiplications, overcoming the limitations of the power trace analysis method in both spatial and temporal dimensions. Following 3DPS preprocessing, we devise the Deep-Learning Side-Channel Attack (DL-SCA) methodologies and successfully recover high-quality input images. In our methodology, we achieve a high accuracy rate with a minimal number of training samples. Furthermore, our work demonstrates the effectiveness of input image recovery under random noise injection. Finally, we delve deeper into the effectiveness of our attack in recovering more intricate images under diverse attack conditions with different loss functions and kernel sizes.