FingerMarks: Robust Multi-Bit Watermarking for Remote Deep Neural Networks

Existing model watermarking methods fail to provide adequate protection for edge intelligence models. This paper innovatively integrates the characteristics of model fingerprinting, proposing a model watermarking method named FingerMarks that enables both model attribution and traceability of edge node users. The method initially constructs a uniform trigger set and an encoding scheme through fingerprint extraction, which effectively distinguishes the host model from independently trained models. Based on the encoding scheme, distinct user IDs are converted and mapped into specific labels, thereby generating distinct watermark-embedded trigger sets. Watermarks are embedded using a progressive adversarial training strategy. Comprehensive evaluation across multiple datasets confirms the method’s performance, uniqueness, and robustness. Experimental results show that FingerMarks effectively identifies the watermarked model while maintaining superior robustness compared to state-of-the-art alternatives.