A Methodology to Evaluate the Security of Block Ciphers Against Neurocryptanalytic Attacks

The increasing development of artificial intelligence has brought a lot of attention to this research area and its potential applications. This paper aims to establish a solid methodology that allows the assessment of the security of a cryptographic algorithm. For our purposes, we chose the neurocryptanalytic approach, which is based on training a neural network to “learn” how to decrypt ciphertexts without knowing the appropriate secret key. We break down the steps, from choosing data sets to defining the appropriate combination of parameters, that will lead us to a solid understanding of how much a cryptographic primitive can resist an attack based on AI techniques. As a target cryptographic algorithm, we opted for a well known and widely used block cipher, which was treated as a black-box (the neural network has no knowledge of its inner operational details). In experiments, our neurocryptoanalytic method was not able to recover a simple message from an AES ciphertext, proving the efficiency of this algorithm.