- category
- watermark-remover
- maturity
- emerging
- approach
- Detection-aware image watermark removal via bidirectional latent optimization: adjusts latent-space representations so the watermark becomes undetectable while bounded changes preserve image quality.
- access
- free
- owns
- An academic research method (arXiv, submitted 2026-10-02) that attacks AI-image source-attribution watermarks by targeting detector rejection rather than merely making the decoded message differ.
- distinctFrom
- Closest are marknull (model-agnostic on-manifold latent decorrelation) and drift-watermark-attack (black-box re-noising and resampling). LiBRA is distinct: a white-box attack that assumes access to the watermark key and decoder, and optimizes bounded changes in a public autoencoder's latent space so average decoding confidence is driven toward random guessing from either direction, avoiding the 'inverted but still detectable' failure of attacks that only push bits away from the original watermark.
- notes
- Submitted 2026-10-02 (authors Saibo Ye et al.). No code or repository link on the arXiv page; access 'free' (paper only). Removal is verified in the paper with an exact two-sided binomial test. Requires watermark key and decoder access, so it is a white-box research attack rather than a blind one. Mechanism details are from the arXiv abstract.
