Research on Copyright Protection Methods for DeepLearning Models Based on Adversarial Examples

The With the widespread application of artificial intelligence technology,copyright protection for deep learning models has become a core industry requirement. To addressthe issue that existing model watermarking schemes require structural modifications to modelswhich may incur potential performance risks, this paper proposes a non-intrusive copyrightprotection method based on adversarial examples. The method leverages the characterizationcapability of adversarial examples regarding model decision boundaries, constructing a sequenceof adversarial samples covering multi-class decision boundaries through targeted adversarial generation algorithms as watermark carriers, while encoding copyright information into the sequencegeneration process. Watermark embedding requires no modification of model parameters orarchitecture, only requiring pre-storage of the adversarial sample set and corresponding outputs asauthentication credentials with authoritative institutions. During copyright verification, ownershipconfirmation can be achieved by comparing the matching degree between the suspected model’sresponse to watermark samples and the authentication credentials. Experiments demonstrate thatthis method achieves considerable copyright verification success rates while preserving originalmodel performance, with adversarial watermark samples showing robustness against commonwatermark removal attacks. This research provides a secure and interpretable novel solution fordeep learning model protection.