Paper 2021/1216

Toward Optimal Deep-Learning Based Side-Channel Attacks: Probability Concentration Inequality Loss and Its Usage

Akira Ito, Rei Ueno, and Naofumi Homma

Abstract

In this paper, we present solutions to some open problems for constructing efficient deep learning-based side-channel attacks (DL-SCAs) through a theoretical analysis. There are two major open problems in DL-SCAs: (i) the effect of the difference in secret key values used for profiling and attack phases is unclear, and (ii) the optimality of the negative log-likelihood (NLL) loss function used in the conventional learning method is unknown. These two problems have hindered the accurate performance evaluation and optimization of DL-SCAs. To address the problem (i), we clarified the strict conditions under which the use of different correct keys in profiling and attack phases affects the performance of DL-SCA. For the problem (ii), we then analyzed the relationship between the NLL loss and direct performance metrics of DL-SCAs (i.e., success rate (SR)/guessing entropy (GE)) and proved that the minimum NLL loss is sufficient but not necessary to achieve the optimal distinguisher of DL-SCA. This explains why DL-SCA succeeds even when the NLL loss is large and motivated us to design a new loss function. Based on the above analysis result, we also propose a new loss function called the probability concentration inequality (PCI) loss function. We derive the PCI loss as an upper bound of GE and a lower bound of the SR using a probability concentration inequality. Minimizing the PCI loss during training can directly optimize the GE and SR of the subsequent attack phase. In this paper, we describe the characteristics of PCI loss and NLL loss and introduce a new learning method that takes full advantage of the characteristics. We also analytically investigate the difference between the PCI loss and ranking loss reported in a previous work for a similar purpose and explain the advantage of PCI loss over the ranking loss. Finally, we validate the analysis and demonstrate the effectiveness of the proposed DL-SCA using the PCI loss through experimental attacks on public datasets.