Protecting Copyrighted Material with Unique Identifiers in Large Language Model Training

원문 초록

A major public concern regarding the training of large language models (LLMs) is whether they abusing copyrighted online text. Previous membership inference methods may be misled by similar examples in vast amounts of training data. Additionally, these methods are often too complex for general users to understand and use, making them centralized, lacking transparency, and trustworthiness. To address these issues, we propose an alternative \textit{insert-and-detection} methodology, advocating that web users and content platforms employ \textbf{\textit{unique identifiers}} for reliable and independent membership inference. Users and platforms can create their own identifiers, embed them in copyrighted text, and independently detect them in future LLMs. As an initial demonstration, we introduce \textit{ghost sentences}, a primitive form of unique identifiers, consisting primarily of passphrases made up of random words. By embedding one ghost sentences in a few copyrighted texts, users can detect its membership using a perplexity test and a \textit{user-friendly} last-$k$ words test. The perplexity test is based on the fact that LLMs trained on natural language should exhibit high perplexity when encountering unnatural passphrases. As the repetition increases, users can leverage the verbatim memorization ability of LLMs to perform a last-$k$ words test by chatting with LLMs without writing any code. Both tests offer rigorous statistical guarantees for membership inference. For LLaMA-13B, a perplexity test on 30 ghost sentences with an average of 7 repetitions in 148K examples yields a 0.891 ROC AUC. For the last-$k$ words test with OpenLLaMA-3B, 11 out of 16 users, with an average of 24 examples each, successfully identify their data from 1.8M examples.