Watermark Forensics for Generative Models: An Information-Theoretic Perspective

What the math gives to ML

The paper supplies a useful decomposition of watermark information into a position-wise conditional mutual-information profile, rather than treating watermark strength as a single scalar. Its key operational insight is that detection, attribution, extraction, and localization consume different resources: attribution among N users requires total information of order log N, while crop-robust localization obeys a footprint-resolution product lower bound. These results can be transferred into trainable watermarking objectives for autoregressive generators by explicitly shaping per-token information and enforcing the unavoidable tradeoff between compact marks and crop robustness. The most practical experiments are to train a keyed generator/decoder with variational estimates of conditional mutual information and compare profile-shaped marks against uniform KL-bias baselines.