Back Propagation Neural Network Based Leakage Characterization for Practical Security Analysis of Cryptographic Implementations

Side-channel attacks have posed serious threats to the physical security of cryptographic implementations. However, the effectiveness of these attacks strongly depends on the accuracy of underlying side-channel leakage characterization. Known leakage characterization models do not always apply into the real scenarios as they are working on some unrealistic assumptions about the leaking devices. In light of this, we propose a back propagation neural network based power leakage characterization attack for cryptographic devices. This attack makes full use of the intrinsic advantage of neural network in profiling non-linear mapping relationship as one basic machine learning tool, transforms the task of leakage profiling into a neural-network-supervised study process. In addition, two new attacks using this model have also been proposed, namely BP-CPA and BP-MIA. In order to justify the validity and accuracy of proposed attacks, we perform a series of experiments and carry out a detailed comparative study of them in multiple scenarios, with twelve typical attacks using mainstream power leakage characterization attacks, the results of which are measured by quantitative metrics such as SR, GE and DL. It has been turned out that BP neural network based power leakage characterization attack can largely improve the effectiveness of the attacks, regardless of the impact of noise and the limited number of power traces. Taking CPA only as one example, BP-CPA is 16.5% better than existing non-linear leakage characterized based attacks with respect to DL, and is 154% better than original CPA.