Watermarking Protects Proprietary Datasets in Generative AI Models

Protecting proprietary datasets used to train generative AI models, particularly large language models, is a significant challenge. Traditional methods for inferring whether specific data points were part of a model's training set have proven difficult. This new work introduces an alternative approach leveraging output watermarking. The core idea is that if a model is trained on partially watermarked data, its outputs will retain a "radioactive" trace of that watermark. By comparing this watermark-based detection method against conventional loss-based membership inference techniques, the study shows that watermarking can achieve similar performance in identifying training data members, especially when a substantial portion of the training dataset has been watermarked. This offers a promising new avenue for data protection and intellectual property enforcement in the AI domain.