Seonglae ChoAI Copyright
DNN Watermarking
AI Watermarks
cryptographic chain-of-custody
Anthropic
Watermarks are used to confirm that content came from a specific model, but editing can weaken the signal, and the absence of a detected watermark is not evidence of human authorship. They can be used as a complement to learning-based detection.
TextGrain
Watermark Basin:Knowledge Distillation watermark
워터마크가 삽입된 데이터로 학습하면 Student 모델의 파라미터 공간에 고유한 방향성인 Watermark Basin이 형성됨. 양자화, Pruning, RL 후속 학습에도 탐지 가능. 다만 추가 SFT나 강한 Paraphrasing은 Black-box 탐지 성능을 크게 떨어뜨림.
