IPNetTool: Watermarking and Chaos for copyright protection of image classification models

Deep neural network (DNN) models have demonstrated significant success in large-scale image datasets, facilitating information exchange over networks for various purposes, including user identification, remote patient health monitoring, early disease detection, and personalized medical treatments. Given the increasing reliance on DNN models for critical applications, ensuring their copyright protection has become a crucial concern in several sensitive domains. This study examines copyright violation issues in DNNs that are subjected to different types of attacks. To address this challenge, we propose a unified framework, IPNetTool, which integrates black-box and white-box watermarking with a chaotic map. This method embeds the generated hash watermark into the trigger images. The watermark is then divided into $k$-chunks. We leverage novel chaos-based techniques to determine the embedding location within the model parameters, after which the chunks are imperceptibly concealed in selected model layers. To evaluate the effectiveness of IPNetTool, we assess the watermarked model in both black-box and white-box scenarios on the receiver’s end. Experimental results demonstrate that IPNetTool enables two-stage ownership verification of four different image classification models under pruning, fine-tuning, and overwriting attacks while reducing the original task accuracy by less than 1% on average. To the best of our knowledge, this is the first attempt to develop a method for the copyright protection of DNN models using chaos-based embedding of a generated hash watermark and two-stage ownership verification via watermarking.