VertMark: A Unified Training-Free Robust Watermarking Framework for Vertical Domain Pre-trained Language Models
VertMark is a new method for protecting Vertical Domain Pre-trained Language Models (VPLMs) from copyright infringement. It creates a training-free watermarking system by establishing hidden semantic links between rare trigger tokens and common domain words. This allows model owners to embed ownership information directly into the model's word embeddings without needing extra training, ensuring the watermark remains robust against attacks like pruning and quantization across medical, finance, and legal domains.