Deep Learning-Based Cryptanalysis of a Simplified AES Cipher

Recently, Deep Neural Networks have shown great deal of reliability and applicability as its applications spread in different areas. This paper proposes a cryptanalysis model based on Deep Neural Network, the neural network takes in plaintexts and their corresponding ciphertexts to predict the secret key of the cipher. We proposes two different approaches, in the first we use multi-layer perceptron (MLP). While in the second, the cryptanalysis problem is modeled as a multi-label classification problem, we introduce appropriate Deep Neural Network based methods for tackling such problem. We illustrate the effectiveness of the approach of the DNN-based cryptanalysis by attacking on Simplified AES block cipher. Therefore, specific metrics are readapted to the cryptanalysis context and used to evaluate the proposed schemes. The results indicate that treating cryptanalysis problem as multi-label classification is more suitable and can be a useful and promising tool for the cryptanalysis task.