Robust LLM Watermarking with Minimal Semantic Distortion for IP Protection
Researchers have developed new methods for watermarking large language models (LLMs) to protect intellectual property and track usage. ArcMark, one new technique, embeds multiple bytes of information into text without altering the LLM's output distribution or perplexity. Another approach, SAFESEAL, uses key-conditioned sampling to preserve semantic fidelity and detect ownership, even against adversarial attacks. TextSeal, a third method, offers localized detection and can transfer its watermark signal through model distillation, making it effective against unauthorized use and replication. AI
IMPACT These watermarking advancements could enable better tracking of LLM-generated content and protect against unauthorized use and distillation.