Robust Watermarking
Overview
Robust watermarking studies how to attach detectable signals to text, images, audio, video, data, and AI models while preserving their quality or utility. Approaches include embedding during generation or in model parameters, using semantic or latent representations, spreading payloads across signals, and adding error-correcting codes or attack-aware training. Evaluation tests transformations such as paraphrasing, compression, cropping, regeneration, fine-tuning, and targeted forgery, with benchmarks increasingly measuring false positives and perceptual quality as well as recovery. The central tension is that stronger signals may distort content, while sophisticated edits can erase or forge marks without obvious damage, so robustness depends on the attack setting and cannot be assumed universally.
Latest papers
207 papers35 papers in the last four weeks, up 40% on the four weeks before. 0.2% of all new papers.