DeepMark: Embedding Watermarks into Deep Neural Network Using Pruning
Chenqi Xie, Ping Yi, Baowen Zhang, Futai Zou
Abstract
With the rapid development of artificial intelligence in recent years, the deep neural network model has been used in many fields such as speech and images due to its excellent performance, and has achieved remarkable results. As we all know, training a deep model requires a lot of time and resources. But these trained deep learning models are very easy to be copied and diffused. Therefore, the protection of intellectual property rights of the model has gradually attracted people’s attention. A series of algorithms or technologies came into being, and one of them is model watermarking technology. Model watermarks can function like digital watermarks. Once the model is stolen, watermarks can prove the copyright of model by verifying the watermarks, maintain its intellectual property rights, and protect the model. This paper proposes a model watermark generation method based on pruning. Where to prune is selected by the calculation result of connection sensitivity, and then the information is embedded by pruning. Compared with the four proposed model watermarking methods, our method has higher fidelity and reliability. Experiments show that our watermarking method is robust against fine-tuning and weight pruning.