GANMarked: Using Secure GAN for Information Hiding in Digital Images

As digital images become increasingly sophisticated, they raise significant security concerns, including the copyright violation, data leakage and identity theft. Deep learning-based data hiding techniques conceals mark within media carriers, enabling both error-free mark extraction and lossless carrier restoration. However, the challenge of enhancing watermark robustness data while ensuring imperceptibility, security, embedding capacity, and model security becomes increasingly pronounced in deep learning environment. In this paper, we present GANMarked, a robust watermarking method embedding a secure mark into the media carriers, based on a generative adversarial network (GAN). First, we utilize an improved autoencoder-based network for secure generation of encoded mark by encoding two individual watermarks into one. Second, the encoded mark imperceptibly embedding into the media carriers using GAN network. Third, the extraction network considers only the marked media as input and robustly recovers the hidden mark at the receiver side. In addition to media security, we fine-tuned the deep watermarking network using secret trigger key to verify the ownership of suspicious models if any piracy or infringements occur. Lastly, decoder network reconstructs the encoded media into the individual one. Our method has been empirically validated across multiple standard datasets, consistently maintaining high imperceptibility, robustness and security, even with variations in hybrid noise during mark extraction. Further, the results demonstrate that the proposed method significantly outperforms other existing methods in terms of imperceptibility and robustness while ensuring reversibility and security.