Practical Persistent Fault Attacks on AES with Instruction Skip

Persistent Fault Attacks (PFA) have emerged as an active research area in embedded cryptography. This attack exploits faults in one or multiple constants stored in memory, typically targeting S-box elements. In the literature, such persistent faults primarily induced by bit flips in storage, often achieved through laser fault injection techniques. In this paper, we demonstrate that persistent faults can also be induced through instruction skips, which can easily be achieved with almost any fault injection methods (e.g., voltage/clock glitching, electromagnetism). Specifically, we target AES implementations that dynamically generate the S-box table at runtime, during the initialization phase, before executing the first AES operation. We illustrate this with an attack on the AES implementation in the MbedTLS library, where a clock glitch is inserted during the S-box generation. Secondly, we introduce, to our knowledge, the first PFA that targets a constant other than the S-box elements. We show that faulting a round constant involved in the AES key schedule is sufficient to recover the key by a differential analysis. Compared to previous PFAs that rely on statistical analysis requiring hundreds to thousands of ciphertexts, our approach needs only three correct-faulty ciphertexts pairs. We showcase this attack with an experiment on the MbedTLS AES implementation, using a clock glitch in the round constant generation.