De-END: Decoder-driven Watermarking Network

With recent advances in machine learning, researchers are now able to solve traditional problems with new solutions. In the area of digital watermarking, deep-learning-based watermarking technique is being extensively studied. Most existing approaches adopt a similar encoder-driven scheme which we name END (Encoder-NoiseLayer-Decoder) architecture. this paper, revamp architecture and creatively design decoder-driven network dubbed De-END greatly outperforms END-based methods. The motivation for designing originated from potential drawback discovered architecture: encoder may embed redundant features that not necessary decoding, limiting performance whole network. We conducted detailed analysis found such limitations caused by unsatisfactory coupling between decoder END. addresses drawbacks adopting Decoder-Encoder-Noiselayer-Decoder De-END, host image firstly processed generate latent feature map instead directly fed into encoder. This concatenated original watermark message then change crucial as it makes shared thus better coupled. extensive experiments results show framework state-of-the-art (SOTA) deep learning both visual quality robustness. On premise same structure, (measured PSNR) improves 1.6dB (45.16dB $\rightarrow$ 46.84dB), extraction accuracy after JPEG compression (QF=50) distortion more than 4% ( notation="LaTeX">$94.9\%\rightarrow 99.1\%$ ).