NeuroCrypt: Entropy-Constrained Neural Key Generation via Contrastive Learning and Fuzzy Extraction From Biometrics

In the evolving field of biometric cryptography, generating high-entropy cryptographic keys directly from noisy physiological signals remains a challenging task. This paper proposes <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">NeuroCrypt</i>, a novel framework for robust and entropy-constrained neural key generation that integrates supervised contrastive learning with fuzzy extraction. Unlike traditional biometric systems that suffer from low entropy or vulnerability to variability, NeuroCrypt learns compact and discriminative embeddings optimized for both reproducibility and randomness. A differentiable quantization layer enables seamless binary key generation, while an entropy-aware loss function ensures uniform bit distribution. The framework further incorporates a fuzzy extractor to tolerate biometric noise and facilitate exact key regeneration. Extensive experiments on fingerprint, iris, and facial biometric datasets demonstrate that NeuroCrypt achieves an average Genuine Acceptance Rate (GAR) of 97.3%, a mean Bit Error Rate (BER) of 3.4%, and an Equal Error Rate (EER) of 3.2%, while generating cryptographic keys with entropy exceeding 250 bits. We note that while BER remains below 3.5% on average, the FaceScrub dataset exhibits a slightly higher BER of 4.2%, highlighting the impact of cross-dataset variability. These results substantially outperform existing neural biometric key generation approaches. Moreover, the system demonstrates resilience against inversion and GAN-based spoofing attacks, with spoofing success rates below 5% across all datasets, while being optimized for edge deployment with inference latency under 50 ms. Collectively, these properties establish NeuroCrypt as a practical and secure solution for real-world biometric-based cryptographic key generation.