A Harmless Owner-Specific Verification Scheme for DNN Copyright Protection
Year: 2026
Type: conference-paper
Abstract: Deep neural networks (DNNs) have achieved significant success in real-world applications. However, protecting their intellectual property (IP) remains extremely challenging. Existing methods include DNN watermarking, fingerprinting, and robust hashing. DNN watermarking is flexible but requires model modifications that reduce model accuracy. Fingerprinting and robust hashing do not alter host models, yet they are more vulnerable and require high computational costs.In this paper, we introduce a novel framework OwnerVerify that enables harmless owner-specific verification. OwnerVerify offers four key features: owner-specific verification, being harmless to the DNN model, a lightweight design, and enhanced robustness. Specifically, OwnerVerify provides ownership verification by recovering specific owner information. Besides, it achieves IP protection for DNNs without altering the original host model, thereby avoiding any reduction in performance. Unlike DNN watermarking and robust hashing methods, it generates secret keys by establishing a stable linkage between flexible owner-specific information and the host model via a gradient descent-based algorithm. These keys are securely stored by a trusted third party and enable a reliable ownership authentication process. Extensive experiments show that OwnerVerify achieves superior security, integrity, and robustness across a wide range of scenarios.
Institutions: Kyushu University, Tohoku University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7216201114