RAE-VWP: A Reversible Adversarial Example-Based Privacy and Copyright Protection Method of Medical Images for Internet of Medical Things

Medical images on the Internet of Medical Things (IoMT) can be easily collected, recognized and analyzed by unauthorized individuals and companies using deep neural networks (DNNs). The illegal use of this data compromises patient privacy and the rights of authorized users. The reversible adversarial example (RAE) is proposed to prevent unauthorized DNNs from recognizing privacy images without affecting authorized users. Nevertheless, existing RAE methods lack research on copyright protection, and unrestricted distribution and sale of these images can seriously jeopardize the copyrights of the image owners. To address this issue, this paper proposes an RAE based on visible watermark perturbation (RAE-VWP). Specifically, we first obtain the watermark embedding region (WER) of the image by the proposed saliency map fusion algorithm and save the information of this region using the reversible data hiding technique. Then, the proposed adaptive memory harmony search algorithm is applied to optimize the position and transparency of the embedded watermark perturbation. Finally, the RAE is generated by embedding the watermark perturbation into the WER of the image using the alpha blending technique. Particularly, the ensemble watermark vaccine is proposed to defend against attacks from watermark removal networks, significantly improving the robustness of RAE-VWP. The original image can be recovered without distortion by extracting the information saved in the RAE. Numerous experiments have shown that RAE-VWP can simultaneously protect image privacy and copyright with excellent reversibility.