Robust Image Watermarking Using Bidirection-Interactive and Context-Aware Networks

Individuals can easily generate highly realistic images using artificial intelligence-generated content technologies, which complicates the verification of images’ ownership rights. This raises potential issues such as spreading misinformation, fraud, and copyright infringement. Digital watermarking is a promising solution to protect the copyright of a digital image by embedding watermarks within it. However, many existing deep learning-based watermarking approaches struggle to simultaneously resist multiple attacks effectively and maintain the quality of watermark images. In this paper, we propose a bidirectional-interactive and context-aware (BICA) deep network designed to enhance the robustness of the watermark while maintaining the quality of the encoded image. We propose a new attention module in the encoder to improve the invisibility and robustness of the watermarked images by implementing an adaptive two-way interaction between local and global features. Additionally, we employ fine-grained downsampling to enhance the attention module’s ability to capture comprehensive feature information. Extensive experimental results demonstrate that the BICA network can embed watermark information into an image without compromising image quality. For instance, BICA has an accuracy exceeding 95% against various moderate noise attacks, with average PSNR and SSIM values of 40.4021 dB and 0.9943, respectively.