Red Team Cryptanalysis of Lattice-Based Post-Quantum Cryptography Using Transformers

The rapid advancement of quantum computing presents a substantial threat to classical cryptographic protocols, accelerating the development of post-quantum cryptography (PQC) schemes, particularly those based on the Learning With Errors (LWE) problem. In parallel, recent progress in artificial intelligence has enabled Transformer-based models-such as SALSA and VERDE-to mount effective cryptanalytic attacks, raising concerns about the resilience of lattice-based PQC. This paper presents a comprehensive empirical evaluation of SALSA (a sequence-to-sequence Transformer with sinusoidal position embeddings) and VERDE (an encoder-only Transformer employing Rotary Position Embeddings) in attacking LWE-based systems.

We assess attack efficacy across diverse experimental settings, including variations in lattice dimension (N), modulus size (log 2 (q)), and secret distribution types (binary, ternary, Gaussian). Experimental results show that VERDE consistently outperforms SALSA, offering up to 30% faster training, higher attack success rates in high-dimensional regimes, and greater adaptability to complex secret distributions. In addition to the empirical analysis, we conduct a targeted risk-benefit assessment for enterprise AI infrastructures, with emphasis on PQC deployment in SAP environments.

We identify critical vulnerabilities, evaluate practical mitigation strategies, and discuss the broader implications for quantum-resilient security architectures. This work advances the understanding of AI-driven cryptanalysis in post-quantum settings and provides actionable insights for strengthening enterprise-level cryptographic defenses.