Answers are grounded in this paper and cite their sources.
Localized image watermarking embeds several messages in different image regions, so composite content can be traced to several users. Watermark Anything (WAM) writes raw 32-bit messages and reports the bit accuracy of the clusters it finds; exact message recovery and the control of false attribution to registered users were not measured. Without retraining WAM, we restricted messages to a 16-bit error-correcting codebook, decoded each detected region by soft maximum-likelihood search over all 65,536 codewords, and accepted an identity only if it was registered and scored above a threshold calibrated on held-out images. In a pre-specified evaluation on 3,000 test images, exact recovery under a horizontal flip with contrast change rose from 0.508 to 0.944. A random codebook of the same size performed identically, indicating the gain comes from codeword separation and soft decoding rather than algebraic code structure. Calibrated attribution reached a correct-attribution rate of 0.802 with false attribution below 0.32\% (upper confidence bound), versus 0.579 for a calibrated bit-tolerance rule. WAM's metric overestimated exact recovery (0.961 versus 0.508), and nearby messages were merged. An attacker knowing the public codebook can impersonate users; with secret codeword assignment no impersonation was observed, but forged regions were attributed to other users.
No takes yet. Share an insight, caveat, or question.
Yun et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: