Improved Differential-ML Distinguisher: Machine Learning Based Generic Extension for Differential Analysis

At CRYPTO 2019, Gohr first proposes a deep learning based differential analysis on round-reduced Speck32/64. Then Yadav etal. present a framework to construct the differential-ML (machine learning) distinguisher by combining the traditional differential distinguisher and the machine learning based differential distinguisher, which breaks the limit of the ML differential distinguisher on the number of attack rounds. However, the results obtained based on this method are not necessarily better than the results gained by traditional analysis. In this paper, we offer three novel greedy strategies (M₁, M₂ and M₃ ) to solve this problem. The strategy M₁ provides better differential-ML distinguishers by considering all combinations of classical differential distinguishers and ML differential distinguishers. And the strategy M₂ uses the best ML differential distinguishers to splice classical differential distinguishers forward, while the strategy M₃ adopts the best classical differential distinguishers to splice ML differential distinguishers. As proof of works, we apply our methods to round-reduced Speck32/64, Speck48/72 and Speck64/96 and get some improved cryptanalysis results. For the construction of differential-ML distinguishers, we can reach 11-round Speck32/64, 14-round Speck48/72 and 18-round Speck64/96 with 2 ²⁷, 2 ⁴⁵, 2 ⁶² data respectively.