Hybrid Neural Networks Assess Cryptographic Vulnerability in Post Quantum Security Transitions

Abstract

The accelerating development of quantum computing fundamentally threatens classical cryptographic infrastructure, compelling organizations worldwide to migrate toward post-quantum cryptography (PQC) algorithms. This transition period, however, introduces compounded vulnerability windows wherein both legacy systems and newly deployed PQC implementations remain susceptible to physical side-channel exploitation. This paper proposes a hybrid neural network (HNN) framework that integrates convolutional neural network (CNN) and long short-term memory (LSTM) architectures to systematically assess cryptographic vulnerabilities during PQC security transitions. The HNN model extracts spatial and temporal leakage patterns from power consumption and electromagnetic traces collected from hardware implementations of PQC candidates, including CRYSTALS-Kyber and CRYSTALS-Dilithium. Experiments conducted on ARM Cortex-M4 microcontrollers running NIST-standardized PQC algorithms demonstrate that the proposed HNN framework achieves a key recovery success rate of 94.7% under masked implementations of Kyber-512 with fewer than 500 attack traces, representing a 17.3% improvement over standalone CNN baselines and 12.8% over standalone LSTM baselines. The framework further produces a vulnerability scoring mechanism that provides actionable guidance for security engineers undertaking cryptographic migration. These findings underscore that neural-network-based evaluation tools must be integral to PQC deployment workflows, ensuring that quantum-resistant algorithms do not inadvertently introduce implementation-level vulnerabilities exploitable by classical adversaries before quantum threats materialize.