A Deep-Learning Approach for Predicting Round Obfuscation in White-Box Block Ciphers

It has been proven that side-channel analysis such as differential computation/fault analysis can break white-box implementations without reverse engineering efforts. In 2020, Sun et al. proposed noisy rounds as a countermeasure to mitigate the side-channel attacks on white-box block ciphers. The principle is to desynchronize the computation traces of cryptographic implementations by introducing several redundant round functions. In this paper, we propose a multi-label classification method and three deep-learning models (CNN, RNN, and CRNN) to predict the locations of the obfuscated rounds. The experimental results show that the obfuscation of noisy rounds also could not be identified by the deep-learning model. However, the RNN is more effective than the CNN and CRNN with fewer time costs. Subsequently, we investigate the influence of specific components such as the key, affine masking, and transformation matrix on round obfuscation recognition. The extended experiments demonstrate that without the transformation matrix, the deep learning models can successfully distinguish the noisy rounds.