A novel non-profiled side channel attack based on multi-output regression neural network

Abstract

Differential deep learning analysis (DDLA) is the first side-channel attack (SCA) evaluation using deep learning (DL) in the non-profiled scenario. However, DDLA requires many training processes to distinguish the correct key. In this paper, we propose an SCA evaluation technique using the multi-output regression neural network, which can simultaneously estimate all key hypotheses in a single training process. Consequently, the performance of non-profiled DL-based SCA can improve significantly. Concretely, two multi-output regression models based on multi-layer perceptron (MOR-MLP) and convolutional neural network (MOR-CNN) are introduced against various SCA protected schemes, such as masking, noise generation, and trace de-synchronization countermeasures. Especially, we first suggest using identity labeling for multi-output regression to determine the trend of the training metric for each hypothesis key in the non-profiled SCA scenario. As a result, the correct key can be distinguished easily. The proposed network is fine-tuned with different variants of the shared layer based on both power consumption and electromagnetic radiation data. The experimental results show that our proposed models using identity label work well on different SCA datasets. Significantly, our approach remarkably outperforms the DDLA model in terms of execution time and success rate. Specifically, the results of evaluating the de-synchronized dataset show that the MOR-CNN model performs the attacks up to 40 times faster than ${\textrm{DDLA}}_{\textrm{CNN}}$. Regarding masking dataset, the MOR-MLP model achieves a higher success rate of at least $30\%$ and $44\%$ in the case of using the fastest ${\textrm{DDLA}}_{\textrm{MLP}}$ model and the original ${\textrm{DDLA}}_{\textrm{MLP}}$ model, respectively.