Side-Channel Analysis on Lattice-Based KEM Using Multi-feature Recognition - The Case Study of Kyber

Kyber, selected as the next-generation standard for key encapsulation mechanism in the third round of the NIST post-quantum cryptography standardization process, has naturally raised concerns regarding its resilience against side-channel analysis and other physical attacks. In this paper, we propose a method for profiling the secret key using multiple features extracted based on a binary plaintext-checking oracle. In addition, we incorporate deep learning into the power analysis attack and propose a convolutional neural network suitable for multi-feature recognition. The experimental results demonstrate that our approach achieves an average key recovery success rate of 64.15% by establishing secret key templates. Compared to single-feature recovery, our approach bypasses the intermediate value recovery process and directly reconstructs the representation of the secret key. Our approach improves the correct key guess rate by 54% compared to single-feature recovery and is robust against invalid attacks caused by errors in single-feature recovery. Our approach was performed against the Kyber768 implementation from \(\texttt{pqm4}\) running on STM32F429 M4-cortex CPU.