Invisible Watermark and Hypothesis Testing for Multi-Ownership Verification Framework of Deep Neural Network

With the rapid development of information technology, artificial intelligence models have become the core asset of tech giants, highlighting the importance of copyright protection. Black-box watermarking has gained attention due to its ability to verify the legality of models remotely without relying on internal model information. The challenges lie in embedding and verifying watermarks without accessing internal model parameters, while ensuring the robustness and stealth of the watermark. This paper focuses on discussing effective ways to protect the copyright of black-box watermarking models and detect and prevent unauthorized user behavior that infringes on model ownership. Therefore, this paper proposes a multi-ownership verification framework based on invisible watermark and hypothesis testing, which mainly consists of two parts: watermark generation and copyright verification. The watermark generation uses a dynamic watermarking method to generate invisible backdoor watermark and verifies copyright through hypothesis testing. This framework has high assurance and effectiveness under various models of CIFAR-10 and MNIST datasets and shows strong robustness in VGG and Resnet model pruning attacks. This research provides new ideas and methods for the protection of model copyrights in deep neural networks.