Split ways: Using GAN watermarking for digital image protection with privacy-preserving split model training

In recent years, the significant importance of digital data in the Industrial Internet of Things (IIoT) is receiving more and more attention, followed by more copyright violation challenges to the transmission and storage of sensitive data. To address this issue, we propose a generative adversarial network (GAN)-based image watermarking in privacy-preserving split model training. In the first stage, we trained our model in split ways without the client sharing raw data to reduce privacy leakage, if any. In the second stage, we designed a GAN-based watermarking embedder and extraction network to imperceptibly embed sensitive information while enhancing robustness. Moreover, the sensitive mark is jointly encrypted and compressed before sending it to the server, thus protecting user confidentiality while reducing the bandwidth and storage demand. We tested our proposed scheme using multiple standard datasets such as div2k, CelebA, and Flickr. The results on the div2k datasets showed that the proposed method surpassed several state-of-the-art methods, with average PSNR and NC increasing by 47.75% and 26.72% respectively. Our joint encryption and compression method also achieved superior performance compared with other methods with an average NPCR and UACI increasing by 18.25% and 16.87% respectively. To the best of our knowledge, we are the first to explore a GAN-based watermarking in split learning ways for digital images.