Researchers have developed a new framework called GeoMark to protect copyright in Embedding-as-a-Service (EaaS) applications. This method addresses the trade-off between robustness, utility, and verifiability in existing watermarking techniques. GeoMark utilizes a natural in-manifold embedding as a watermark target and creates geometry-separated anchors, enabling watermark injection within specific local neighborhoods for localized triggering and centralized attribution. Experiments demonstrate that GeoMark maintains downstream utility and geometric fidelity while offering robust copyright verification against various attacks, including paraphrasing and dimensional perturbation, with improved stability and a low risk of false positives. AI
IMPACT This framework could improve the security and integrity of AI services that rely on embeddings, potentially reducing instances of model theft and copyright infringement.
RANK_REASON The cluster contains an academic paper detailing a new technical framework. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
