Micropower: Micro Neural Networks for Side-Channel Attacks

Side-channel attacks recover encryption keys from a target by analyzing power consumption. Deep learning sidechannel attacks require less pre-processing and can defeat countermeasures compared to traditional methods. However, neural networks utilized in side-channel attacks are often complex, which requires substantial storage and memory. In this paper, we propose an iterative pruning algorithm, named MicroPower, which can significantly reduce the size and memory usage of neural networks for side-channel attacks. Our algorithm automatically adjusts pruning parameters of the next iteration based on the results of its previous iteration to derive extremely small neural networks that can still recover keys successfully. We conduct comprehensive evaluation over three existing datasets and demonstrate the effectiveness of our algorithm. For instance, our algorithm can remove 99.86 % of parameters in a CNN and the pruned CNN can still recover keys over ASCADv1 dataset with only 321 traces. Given ASCADv2 dataset, we can remove 91.74 % of parameters in a ResNet and reveal encryption keys with only 519 traces. Our pruning is also compatible with quantization, which can further reduce the size of a neural network. Moreover, we demonstrate that our pruned neural networks can run efficiently and effectively on embedded devices, including Nvidia Jetsons and AMD/Xilinx ZCU104 FPGA boards.