Empirical Hardness of AES Cipher

Kanat Alimanov, Martin Lukac, Shinobu Nagayama


The AES symmetric cipher is widely used as a standard encryption method in various network protocols. Although it has proven resistant to most direct attacks, it hasn’t been extensively studied from the perspective of modern neurocryptanalysis and big data. Therefore, this paper aims to comprehensively analyze the components of the AES protocol, empirically determine their learnability, and provide empirical results that demonstrate the hardness of the cipher. Our analysis involves evaluating the ability of various models to learn the different components of the AES cryptosystem, as well as their combinations. Furthermore, we examine the overall ability of these models to recover a network that estimates the operation of adding the secret key to the input data. Through our research,
we show that AES is indeed resistant to machine learning attacks. However, under certain configurations of the data and the AES cipher, it is possible to recover the decrypting network.


AES; Machine Learning; Neural Networks

Full Text:



  • There are currently no refbacks.