Search

Watermarking is a technique that involves embedding nearly unnoticeable statistical signals within generated content to help trace its source. This work focuses on a scenario where an untrusted third-party user sends prompts to a trusted language model (LLM) provider, who then generates a text from their LLM with a watermark. This setup makes it possible for a detector to later identify the source of the text if the user publishes it. The user can modify the generated text by substitutions, insertions, or deletions. Our objective is to develop a statistical method to detect if a published text is LLM-generated from the perspective of a detector. We further propose a methodology to segment the published text into watermarked and non-watermarked sub-strings. The proposed approach is built upon randomization tests and change point detection techniques. We demonstrate that our method ensures Type I and Type II error control and can accurately identify watermarked sub-strings by finding the corresponding change point locations. To validate our technique, we apply it to texts generated by several language models with prompts extracted from Google's C4 dataset and obtain encouraging numerical results. We release all code publicly at https://github.com/doccstat/llm-watermark-cpd., Comment: 25 pages, 12 figures, 2 tables, NeurIPS 2024

Change point analysis is concerned with detecting and locating structure breaks in the underlying model of a sequence of observations ordered by time, space or other variables. A widely adopted approach for change point analysis is to minimize an objective function with a penalty term on the number of change points. This framework includes several well-established procedures, such as the penalized log-likelihood using the (modified) Bayesian information criterion (BIC) or the minimum description length (MDL). The resulting optimization problem can be solved in polynomial time by dynamic programming or its improved version, such as the Pruned Exact Linear Time (PELT) algorithm (Killick, Fearnhead, and Eckley 2012). However, existing computational methods often suffer from two primary limitations: (1) methods based on direct implementation of dynamic programming or PELT are often time-consuming for long data sequences due to repeated computation of the cost value over different segments of the data sequence; (2) state-of-the-art R packages do not provide enough flexibility for users to handle different change point settings and models. In this work, we present the fastcpd package, aiming to provide an efficient and versatile framework for change point detection in several commonly encountered settings. The core of our algorithm is built upon PELT and the sequential gradient descent method recently proposed by Zhang and Dawn (2023). We illustrate the usage of the fastcpd package through several examples, including mean/variance changes in a (multivariate) Gaussian sequence, parameter changes in regression models, structural breaks in ARMA/GARCH/VAR models, and changes in user-specified models., Comment: 53 pages, 16 figures

Modeling heterogeneity on heavy-tailed distributions under a regression framework is challenging, and classical statistical methodologies usually place conditions on the distribution models to facilitate the learning procedure. However, these conditions are likely to overlook the complex dependence structure between the heaviness of tails and the covariates. Moreover, data sparsity on tail regions also makes the inference method less stable, leading to largely biased estimates for extreme-related quantities. This paper proposes a gradient boosting algorithm to estimate a functional extreme value index with heterogeneous extremes. Our proposed algorithm is a data-driven procedure that captures complex and dynamic structures in tail distributions. We also conduct extensive simulation studies to show the prediction accuracy of the proposed algorithm. In addition, we apply our method to a real-world data set to illustrate the state-dependent and time-varying properties of heavy-tail phenomena in the financial industry., Comment: 23 pages, 9 figures