Deep Learning-Assisted Differential Cryptanalysis on Round-reduced Block Ciphers: What are the advantages?

Abstract

This thesis investigates and evaluates the application of deep learning techniques in differential cryptanalysis of lightweight block ciphers. The main two focuses are on comparison between a hybrid attack pipeline and a classical attack pipeline, and on transfer learning compared to training a model from scratch. The study considers both SPN ciphers and the SIMON32/64 cipher, which is part of the ISO standard for RFID systems. Neural distinguishers based on convolutional neural networks are trained to classify pairs of ciphertexts, as either random noise or ciphertexts produced by the cipher. For the SPN cipher it is integrated into a key recovery attack pipeline. For SIMON32/64 it is compared to the outcome of transfer learning of a pre-trained network. The results show that a hybrid approach is comparable to a classical approach in terms of key recovery attack. The transfer learning enables faster convergence, but does not reach the same accuracy in classification compared to training the model from scratch. These findings contribute to the understanding of the role of machine learning in cryptanalysis, and how it can be further studied to potentially be more useful in the future, and what the security impacts might be for real-world use in especially supply chains using RFID technology.