Paper 2026/734

Assessing Geometric Security of AES Neural Realizations: Linear-Time Key Recovery via Neural Leakage

Abstract

We investigate the security of AES-128/192/256 when implemented as ReLU-based neural networks via the natural sum-of-corners construction. Although these implementations are bit-exact on Boolean inputs, they extend AES into a continuous piecewise-linear function over $\mathbb{R}^{128}$. We show that under real-valued oracle access, such neural realizations admit deterministic linear-time master-key recovery. The attack exploits a geometric property of the natural XOR (AddRoundKey) layer: for corner parameter c < 1, ReLU activations partition the input space into key-dependent linear regions. Using symmetric perturbations, exactly one key hypothesis preserves linear-region membership, enabling bitwise recovery through simple output-equality tests. We formalize this phenomenon via a local separability lemma and obtain attack complexity O(128R) neural queries for R rounds. Experiments over 1,000 independent random keys for AES-128, AES-192, and AES-256 achieve 100% recovery success. The vulnerability is independent of key size and round count; it stems solely from exposing a key-dependent XOR layer inside a continuous neural architecture. Our results reveal a fundamental gap between Boolean security and geometric security in neural implementations of cryptographic primitives.