Deep Learning Side-Channel Attacks Against Lightweight SCA Countermeasure RSM-AES

Recently, deep learning side-channel attacks (DL-SCA) have been actively studied. The attacks against a software-implemented AES with lightweight countermeasures such as misalignment and masking have already been successfully demonstrated by utilizing simple DL-SCA. On the other hand, it is not clarified how much performance is exhibited on hardware-implemented AES circuits. The XORed model using register transition has been used for simple DL-SCA against hardware-implemented AES circuits. However, partial keys on 0, 4, 8, and 12 bytes cannot be revealed by simply applying the XORed model because of data imbalance problems. In the previous study, the ‘mix XORed model’ method had been proposed, and partial keys on 0, 4, 8, and 12 bytes are successfully attacked by utilizing training data of other key bytes. This method can reveal all partial keys on unprotected AES hardware; however, it is difficult to reveal all keys on RSM (Rotating S-Box Masking) AES hardware even with 100,000 traces. In this paper, we experimentally applied the ‘random XORed model’ where waveforms during processing with random keys are used as the training dataset in order to resolve the data imbalance problem on partial keys on 0, 4, 8, and 12 bytes. In our experiments, we demonstrated that all partial keys are successfully revealed against unprotected AES with 600 traces and RSM-AES circuits with 45,000 traces.