International Association
for Cryptologic Research

In conventional deep-learning-based side-channel attacks (DL-SCAs), an attacker trains a model by updating parameters to minimize the negative log-likelihood (NLL) loss function. Although a decrease in NLL improves DL-SCA performance, the reasons for this improvement remain unclear because of the lack of a formal analysis. To address this open problem, this paper explores the relationship between NLL and the attack success rate (SR) and conducts an information-theoretical analysis of DL-SCAs with an NLL loss function to solve open problems in DL-SCA. To this end, we introduce a communication channel for DL-SCAs and derive an inequality that links model outputs to the SR. Our inequality states that mutual information between the model output and intermediate value, which is named the latent perceived information (LPI), provides an upper bound of the SR of a DL-SCA with a trained neural network. Subsequently, we examine the conjecture by Ito et al. on the relationship between the effective perceived information (EPI) and SR and clarify its valid conditions from the perspective of LPI. Our analysis results reveal that a decrease in NLL correlated with an increase in LPI, which indicates that the model capability to extract intermediate value information from traces is enhanced. In addition, the results indicate that the LPI bounds the SR from above, and a higher upper bound of the SR could directly improve the SR if the selection function satisfies certain conditions, such as bijectivity. Finally, these theoretical insights are validated through attack experiments on neural network models applied to AES software and hardware implementations with masking countermeasures.