A Transfer Learning Approach for Electromagnetic Side-channel Attack and Evaluation

Side channel analysis has posed a huge threat to cryptographic function security. And some mitigation techniques including masking and trace data desynchronization are used to prevent the adversary from attacking. In this article, we propose a transfer learning framework to attack this countermeasure and recover the secret key effectively and efficiently, which can not only reduce the trace number required for side channel attacks but also reduce the time to train deep learning models. We first deserve a well trained deep learning model by taking the simplest source trace synchronization dataset as the source domain. Then we migrate and adjust this well trained model to the target trace desynchronization dataset by pre-training, freezing, fine-tuning, etc. In addition, we implement four different deep learning models including AutoEncoder, CNN, MLP and ResNet into the transfer learning framework for fixed key scenario and variable key scenario to demonstrate the framework scalability and generality. Finally, we evaluate the transfer learning framework using the public ASCAD datasets. The experimental results show that our method can successfully recover the key with 75 traces on the ASCAD synchronous dataset, and the transfer learning framework can solve the situation that the signal is biased, it only needs 60 traces to recover the keys on the random offset 50 dataset, and only 75 to recover the keys on the random offset 100 dataset.