Robust Watermark Imaging via Graph-signal Optimization

The topic of robust watermarking has long been a subject of interest in the field of information hiding, and in these days, the robustness against cross-media attacks, such as screen-capture and printer-capture attacks, have become an increasingly pressing concern. Previous work in the field, such as StegaStamp, one of the most popular open-source screen-shoot resilient watermark scheme, has achieved promising results in terms of decoding accuracy. However, the visual quality of some watermarked images by this scheme is still not inadequate for practical purposes. In this paper, we propose an optimization that could significantly enhances the visual quality of watermarked images. We use a graph-based method to build graph from image blocks, compute a loss designed with inspiration from GLR, and add it to the existing loss function of the neural network during training. As demonstrated by the experiment results, we have achieved substantial improvements in the quality of watermarked images, in both visual quality and metrics, while still maintaining strong robustness and decoding capability.