With Internet technology advancing, artificial intelligence models have become increasingly vital across various industries. However, developing a sophisticated artificial intelligence model demands substantial investment. So how to protect the copyright of these models has become a crucial research focus. Existing methods, such as watermarking and fingerprinting, have shown potential in safeguarding these copyrights. However, these approaches depend on third-party organizations for copyright management. If these organizations encounter attacks and issues, they will fail to provide reliable copyright information. Based on this, inspired by blockchain mechanisms, we propose Copyright GuardChain: protecting intellectual property of deep neural networks via a new copyright blockchain. This framework comprises three core modules: model fingerprint extraction, copyright resource uploading, query and determination. By implementing our fine-grained model feature extraction, designing a hybrid approach combining MinHashLSH and Merkle Tree for model fingerprint uploading, and finally adopting the high confidence similarity to measure the model infringement, we achieve efficient and reliable DNN model copyright protection. Experimental results demonstrate that our approach achieves 93.75% accuracy against various attacks. Moreover, the Copyright GuardChain offers fast on-chain storage, querying, and infringement detection, with a low overhead ranging from 0.61% to 11.41%.