International Association
for Cryptologic Research

CRYSTALS-Kyber, one of the NIST PQC standardization schemes, has garnered considerable attention from researchers in recent years for its side-channel security. Various targets have been explored in previous studies; however, research on extracting secret information from stack-optimized implementations targeting the Cortex-M4 remains scarce, primarily due to the lack of memory access operations, which increases the difficulty of attacks.This paper shifts the focus to the leakage of multiply instructions and present a novel cycle-level regression-based leakage model for the following attacks. We target the polynomial multiplications in decryption process of the stack-optimized implementation targeting the Cortex-M4, and propose two regression-based profiled attacks leveraging known ciphertext and chosen ciphertext methodologies to recover the secret coefficients individually. The later one can also be extended to the protected implementation.Our practical evaluation, conducted on the stack-optimized Kyber-768 implementation from the pqm4 repository, demonstrates the effectiveness of the proposed attacks. Focusing on the leakage from the pair-pointwise multiplication, specifically the macro doublebasemul_frombytes_asm, we successfully recover all secret coefficients with a success rate exceeding 95% using a modest number of traces for each attack. This research underscores the potential vulnerabilities in PQC implementations against side-channel attacks and contributes to the ongoing discourse on the physical security of cryptographic algorithms.

Since Kocher (CRYPTO’96) proposed timing attack, side channel analysis (SCA) has shown great potential to break cryptosystems via physical leakage. Recently, deep learning techniques are widely used in SCA and show equivalent and even better performance compared to traditional methods. However, it remains unknown why and when deep learning techniques are effective and efficient for SCA. Masure et al. (IACR TCHES 2020(1):348–375) illustrated that deep learning paradigm is suitable for evaluating implementations against SCA from a worst-case scenario point of view, yet their work is limited to balanced data and a specific loss function. Besides, deep learning metrics are not consistent with side channel metrics. In most cases, they are deceptive in foreseeing the feasibility and complexity of mounting a successful attack, especially for imbalanced data. To mitigate the gap between deep learning metrics and side channel metrics, we propose a novel Cross Entropy Ratio (CER) metric to evaluate the performance of deep learning models for SCA. CER is closely related to traditional side channel metrics Guessing Entropy (GE) and Success Rate (SR) and fits to deep learning scenario. Besides, we show that it works stably while deep learning metrics such as accuracy becomes rather unreliable when the training data tends to be imbalanced. However, estimating CER can be done as easy as natural metrics in deep learning algorithms with low computational complexity. Furthermore, we adapt CER metric to a new kind of loss function, namely CER loss function, designed specifically for deep learning in side channel scenario. In this way, we link directly the SCA objective to deep learning optimization. Our experiments on several datasets show that, for SCA with imbalanced data, CER loss function outperforms Cross Entropy loss function in various conditions.