| With the continuous application of deep learning in image processing,deep image processing networks have been widely used in various fields such as image classification,object detection,image segmentation,image generation,and image super-resolution.Training an excellent deep image processing network is not easy,as it requires massive labeled data resources,as well as significant computational resources and manual efforts.As a digital product,once the deep image processing model is trained and published,it is vulnerable to illegal copying,redistribution,or abuse by malicious users,posing a significant threat to model copyright.Existing model copyright protection methods mostly focus on classification networks,while there is a scarcity of protection methods for image processing networks with complex input-output and model structures,resulting in a lack of effective copyright protection methods for image processing networks.Moreover,there are still shortcomings in existing image processing network protection methods,such as reduced accuracy after watermark embedding and increased training difficulty due to increased parameters.To address these issues,this paper proposes two copyright protection methods for image processing networks,as follows:(1)To address the problems of manually designed trigger samples and increased parameters after watermark embedding in existing model copyright protection methods,this paper proposes a dual-watermark end-to-end copyright protection framework for deep image processing networks.The trigger sample watermark is automatically generated based on the trigger sample method of adversarial generative networks,which solves the problem of poor transferability of manually designed trigger samples.The model watermark utilizes the proposed embedding algorithm to automatically select the embedding position,which effectively resists forgery attacks without increasing model parameters.Experimental results on deep image processing networks such as UNET,UNET++,and FCN demonstrate that the proposed framework has good performance in fidelity,integrity,and robustness,and the embedding process maintains good network segmentation accuracy,without increasing the number of network parameters.(2)To address the problems of reduced accuracy caused by modifying model parameters and inadequate active protection in most existing methods,this paper proposes a model copyright protection method based on gate structures and feature selection.Firstly,a gate structure is ingeniously designed by combining the authorization code to control the model’s usage.When the correct authorization code is entered,the model is normally used;otherwise,the model accuracy is significantly reduced,preventing the model from being used,achieving the goal of actively protecting the model’s copyright.Secondly,the watermark is embedded into the unimportant features of the model through feature selection,which can be extracted from the model even if it is maliciously distributed,thus verifying the model’s copyright.Experiments show that the proposed method has good performance in both classification networks and deep image processing networks and can actively protect the model and verify its copyright after infringement. |
