February 21, 2024
This paper investigates the radioactivity of LLM-generated texts, i.e. whether it is possible to detect that such input was used as training data. Conventional methods like membership inference can carry out this detection with some level of accuracy. We show that watermarked training data leaves traces easier to detect and much more reliable than membership inference. We link the contamination level to the watermark robustness, its proportion in the training set, and the fine-tuning process. We notably demonstrate that training on watermarked synthetic instructions can be detected with high confidence (p-value < 10−5) even when as little as 5% of training text is watermarked. Thus, LLM watermarking, originally designed for detecting machine-generated text, gives the ability to easily identify if the outputs of a watermarked LLM were used to fine-tune another LLM.
Publisher
ARXIV
July 23, 2024
Llama team
July 23, 2024
June 25, 2024
Elena Voita, Javier Ferrando Monsonis, Christoforos Nalmpantis
June 25, 2024
June 25, 2024
Min-Jae Hwang, Ilia Kulikov, Benjamin Peloquin, Hongyu Gong, Peng-Jen Chen, Ann Lee
June 25, 2024
June 14, 2024
Sheng-Chieh Lin, Akari Asai, Minghan Li, Barlas Oguz, Jimmy Lin, Scott Yih, Xilun Chen
June 14, 2024
Product experiences
Foundational models
Product experiences
Latest news
Foundational models