Using watermarks to remove watermarks from images

Image watermarking is supposed to protect copyright and prevent deepfake abuse, but this work reveals a fundamental vulnerability: watermarks can be weaponized to attack watermarks. By treating watermark removal as an analogue to watermarking itself, the authors show that simply applying a second watermark to an already-watermarked image effectively erases the first one. They also develop a classifier that identifies existing watermarks with 87.8–95.3% accuracy across datasets. Together, watermark detection plus re-watermarking forms a generic, black-box attack requiring no gradient information or model knowledge—and outperforms most existing sophisticated removal methods.