Paper 2025/862 Original: eprint.iacr.org/2025/862 ↗

Distinguishing Full-Round AES-256 in a Ciphertext-Only Setting via Hybrid Statistical Learning

Abstract

The security of block ciphers such as AES-128, AES-192, and AES-256 relies on the assumption that their ciphertext outputs are computationally indistinguishable from random permutations. While distinguishers have been proposed for reduced-round variants or under non-standard models such as known-key or chosen-key settings, no effective distinguisher has been demonstrated for the full-round AES ciphers in the standard secret-key model. This work introduces FESLA (Feature Enhanced Statistical Learning Attack), a hybrid statistical learning framework that integrates outputs from a suite of classical statistical tests with machine learning and deep learning classifiers to construct ciphertext-only distinguishers for AES-128, AES-192, and AES-256. In contrast to existing approaches based on handcrafted or bitwise features, FESLA aggregates intermediate statistical metrics as features, enabling the capture of persistent structural biases in ciphertext distributions. Experimental evaluation across multiple datasets demonstrates consistent 100% classification accuracy using Support Vector Machines, Random Forests, Multi-Layer Perceptron, Logistic Regression, and Naïve Bayes classifiers. Generalization and robustness are confirmed through k-fold cross-validation, including on previously unseen ciphertext samples. These results establish the first ciphertext-only distinguishers for full-round AES-128, AES-192, and AES-256 under the secret-key model, and underscore the potential of machine learning–augmented cryptanalysis based on principled statistical feature engineering.