Multi-Byte Power Analysis: A Generic Approach Based on Linear Regression

Abstract

Linear regression used to be known as a robust side-channel analysis (SCA) method as it makes use of independent bits leakage. This leakage assumption is more general than Hamming weight/Hamming distance model used in correlation power analysis (CPA). We find that in many common scenarios, linear regression is not only an alternative but also a more efficient tool compared with CPA. This paper proposes a generic SCA approach based on linear regression called multi-byte power analysis (MPA) that can be applied to any number of bytes instead of one single byte when performing SCA. Two typical cases are illustrated in this paper. One is recovering keys with XOR operation leakage and the other one is chosen plaintext attack on block ciphers with leakages from round output. Simulation results are given to compare with traditional CPA in both cases. MPA achieves up to 400% and 300% improvements for the corresponding case compared with CPA, respectively. Experiments with AES on SAKURA-G board also prove the efficiency of MPA in practice, where 128 key bits are recovered with 1500 traces using XOR operation leakage and one key byte is recovered with only 50 chosen-plaintext traces in the other case.