Deep Learning-based Reliability Control for Robust Image Watermarking in Encrypted and Compressed Domains

Abstract

In order to meet the increasing need for cloud storage and multimedia transmission, ownership authentication of digital images must be securely achieved. Traditional water-marking techniques may lose robustness after image encryption and lossy compression. In this paper, a deep learning-based reliability control framework of image watermarking in the encrypted and compressed domains, using a Recurrent Neural Network (RNN) structure, is presented. In this approach, the RNN structure can effectively model the sequential dependencies among image blocks to adaptively adjust the watermark strength according to the contextual features, thereby achieving a balance between robustness and imperceptibility. The proposed method can be securely applied in the encrypted domain and could resist lossy compression attacks, including JPEG and HEVC intra-coding attacks. In the experimental results, improvements in terms of PSNR, SSIM, and BER are achieved compared to traditional methods. The RNN-based reliability control framework of image watermarking is efficient and secure for real-time image and video applications.