doi: 10.7763/IJCTE.2026.V18.1405
A Fully Trained Neural Network Surrogate of the AES128 Encryption
Mohamed El Hassani*
,
Mahdi Madani ,
El-Bay Bourennane
- ImViA Laboratory, Université Bourgogne Europe, IMVIA UR 7535 Dijon, France
- Manuscript receivedMay 31, 2026
- revisedJune 24, 2026
- acceptedAugust 28, 2026
- publishedOctober 19, 2026
Abstract
This paper presents a fully learned neural network surrogate of the Advanced Encryption Standard (AES) encryption algorithm under any oracle-provided key. Each round is split into three learned modules—ShiftRows-SubBytes, MixColumns, and AddRoundKey—with the round key given as input so one model handles arbitrary keys. Trained on fewer than 50,000 samples, the surrogate empirically attains perfect bit accuracy with zero observed errors over millions of unseen evaluations, without overfitting. Inspecting the models reveals behaviour atypical of deep learning: accuracy drops with depth, hinges on a sharp hidden-dimension threshold, and rests on near-binary sparse weights that collapse under pruning. As an exact, differentiable model of AES, the surrogate provides an empirical baseline for gradient-based and surrogate-assisted analysis of cryptographic primitives.
Keywords
- deep learning
- Advanced Encryption Standard (AES)
- surrogate modeling
- gradient-based learning
- block cipher emulation
How to Cite
Mohamed El Hassani, Mahdi Madani, and El-Bay Bourennane, "A Fully Trained Neural Network Surrogate of the AES128 Encryption," International Journal of Computer Theory and Engineering, vol. 18, no. 4, pp. 253-259, 2026. https://doi.org/10.7763/IJCTE.2026.V18.1405
Copyright & License
Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).