NNFMAC: A Neural Network Fingerprinting-Based Model Authentication Code Scheme

As deep learning–based AI proliferates, model theft and plagiarism pose increasing Intellectual Property (IP) risks. However, watermarking alters model weights and can degrade performance, while fingerprinting often merely verifies uniqueness or requires heavy computation. In this article, we propose a Neural Network Fingerprinting-Based Model Authentication Code (NNFMAC) scheme that verifies both model uniqueness and ownership without affecting performance. NNFMAC extracts key weights from a trained model, applies a median-based method to generate a unique binary fingerprint, and uses this fingerprint as a codebook to encode ownership information via a newly designed index-based function with expansion, producing reliable authentication codes. This non-intrusive approach integrates fingerprinting for uniqueness verification and authentication coding for ownership verification, delivering comprehensive model IP protection while preserving the model’s original performance. Extensive experiments demonstrate that NNFMAC preserves model accuracy without additional training overhead, unlike other watermarking schemes that degrade accuracy by 0.36–1.53%. It achieves bit error rates of 0.12 under weight perturbation, 0.03 under fine-tuning, 0.08 under pruning, and 0.09 under weight shifting attacks, which are substantially lower than the 0.51, 0.49, 0.46, and 0.22 reported in prior work, while consistently outperforming state-of-the-art schemes in effectiveness, efficiency, and robustness.