A Statistical Hypothesis Testing Framework for Data Misappropriation Detection in Large Language Models
This addresses privacy and legal concerns for LLM developers and copyright holders by providing a detection method for data misappropriation, though it is incremental as it builds on existing watermarking and statistical testing concepts.
The authors tackled the problem of detecting whether a large language model (LLM) has incorporated data generated by another LLM, by proposing a statistical hypothesis testing framework that embeds watermarks into copyrighted training data, with explicit control of type I and type II errors and demonstrated empirical effectiveness in experiments.
Large Language Models (LLMs) are rapidly gaining enormous popularity in recent years. However, the training of LLMs has raised significant privacy and legal concerns, particularly regarding the distillation and inclusion of copyrighted materials in their training data without proper attribution or licensing, an issue that falls under the broader concern of data misappropriation. In this article, we focus on a specific problem of data misappropriation detection, namely, to determine whether a given LLM has incorporated the data generated by another LLM. We propose embedding watermarks into the copyrighted training data and formulating the detection of data misappropriation as a hypothesis testing problem. We develop a general statistical testing framework, construct test statistics, determine optimal rejection thresholds, and explicitly control type I and type II errors. Furthermore, we establish the asymptotic optimality properties of the proposed tests, and demonstrate the empirical effectiveness through intensive numerical experiments.