Side-Channel Attack on Deep Reinforcement Learning of AES
Pei Peng, Meiling Zhang, Dong Zheng · 2022 4th International Conference on Natural Language Processing (ICNLP) · 2022
In the research field of side-channel attacks, label-based modeling attacks based on machine learning have been successfully applied to block cipher AES. Compared with non-modeling attacks, modeling attacks have a large workload of data collection and processing in the early stage, but their advantage lies in the high reusability of the model. Once the model is established successfully, the key can be recovered with less traces. First, implement the AES algorithm on the Chipwhisperer platform and collect the energy traces of the process, borrow the CPA technology to extract the points of interest, and the obtained points of interest are the data set for model training. This paper applies deep reinforcement learning to the research of symmetric cipher AES, and realizes a label-free training model. Experiments show that deep reinforcement learning technology can be successfully applied to side-channel attacks. The model can predict that the correct rate of both synchronous and asynchronous energy trajectories is as high as 91% or more. This data shows that the model can attack the correct key with an energy trace.