Toward Practical Deep Blind Watermarking for Traitor Tracing

초록

Traitor tracing via blind watermarking is a promising solution to protect copyright. Recently, it was demonstrated that deep learning-based blind watermarking methods could outperform traditional watermarking methods in terms of robustness against distortions. However, we found they could sacrifice the imperceptibility as a trade-off. To handle this challenge, we propose a novel method consisting of a deep blind model and watermarking strategy. For the purpose, we first investigate the fundamental components of the basic deep blind watermarking model, and empirically show how the performance changes when each component is modified with respect to the robustness. Based on it, we construct a deep blind watermarking encoder, CFC+CONCAT, which can encode watermarks in a robust way against distortions without imperceptibility degradation. We then propose a watermarking strategy to make deep blind watermarking robust by increasing watermarking capacity (by splitting a large image into small patches), and using the effect of distortion types on the robustness we found. According to the experiments, our method achieved 5.46 higher PSNR on average than the baseline methods with comparable robustness under various distortions when watermarking in the 3 x 256 x 256 image. Also, the training time and VRAM usage are reduced by less than 1/4, demonstrating our method can mitigate the trade-off between robustness and imperceptibility, and achieve lightweight training.

키워드