Paper 2022/1468

Vulnerability Assessment of Ciphers To Fault Attacks Using Reinforcement Learning

Abstract

A fault attack (FA) is one of the most potent threats to cryptographic applications. Implementing a FA-protected block cipher requires knowledge of the exploitable fault space of the underlying crypto algorithm. The discovery of exploitable faults is a challenging problem that demands human expertise and time. Current practice is to rely on certain predefined fault models. However, the applicability of such fault models varies among ciphers. Prior work discovers such exploitable fault models individually for each cipher at the expanse of a large amount of human effort. Our work completely replaces human effort by using reinforcement learning (RL) over the huge fault space of a block cipher to discover the effective fault models automatically. Validation on an AES block cipher demonstrates that our approach can automatically discover the effective fault models within a few hours, outperforming prior work, which requires days of manual analysis. The proposed approach also reveals vulnerabilities in the existing FA-protected block ciphers and initiates an end-to-end vulnerability assessment flow.