Chameleon DNN Watermarking: Dynamically Public Model Ownership Verification

Wei Li, Xiaoyu Zhang, Shen Lin, Xinbo Ban, Xiaofeng Chen

Abstract: Deep neural network (DNN) has made unprecedented leaps in functionality and usefulness in the past few years, revolutionizing various promising fields such as image recognition and machine translation. The trainer’s high-performance DNNs are often considered intellectual property (IP) due to their expensive training costs. However, one pre-trained model may face various infringement problems when hacked by a malicious user, such as illegal copying or secondary selling. Digital watermarking is one of the effective methods currently used for model ownership verification. Nonetheless, limited by the ex-ante nature of the watermark embedding phase and the ex-post nature of the verification phase, previous research has only supported private verification or one-time public verification, failing to achieve multiple public verifications. In this paper, we introduce the definition of chameleon DNN watermarking and propose the first DNN watermarking scheme based on chameleon commitment, which allows multiple public verifications to declare the owner’s model ownership without exposing the core watermark information. We give a comprehensive security analysis of the verification scheme of chameleon DNN watermarking and prove by experiments that chameleon DNN watermarking can maintain the high-performance and robustness of the model.