DESIGN OF SYMMETRIC CRYPTOGRAPHIC DIFFERENTIAL DISTINGUISHER BASED ON DEEP LEARNING

Research in the field of cryptanalysis demonstrates that differential discriminators based on neural networks significantly outperform traditional methods in identifying weaknesses in simple encryption algorithms. This is due to their ability to detect complex patterns in data that may go unnoticed by classical approaches. However, despite their high performance, there are limitations in terms of the accuracy of the discrimination and the maximum number of rounds that can be compromised for certain ciphers. The purpose of this study is to overcome these shortcomings by optimizing the architecture of the neural network (NN) and the structure of the input data, according to the Gohr model. As part of this work, several key components of the differential discriminator were improved: the convolutional module, the residual module, and the prediction module. The optimization of these components allowed to significantly increase the efficiency in recognizing differential patterns in ciphertexts. In addition, special attention was paid to optimizing the structure of the input data, which made it possible to more accurately identify the characteristics of the ciphertext and information abou