Statistics and Probability seminar
Speaker: Soham Bonnerjee, University of Pennsylvania
Host: Debraj Das
Title: Fast Watermark Segmentation in LLM generated texts (and other application of epidemic change-points)
Time, day and date: 5:00:00 PM - 6:00:00 PM, Tuesday, August 18
Venue: Ramanujan Hall
Abstract: Epidemic change-points has seldom received attention in statistics community; however, it has  important, and somewhat surprising applications in critically important and modern problems. In this talk, we primarily focus on the "watermark segmentation" problem. With the growing use of large language models, concerns over content authenticity have spurred a variety of watermarking schemes. These schemes use secret keys to detect machine-generated text while remaining imperceptible to readers. Detection typically reduces statistical hypothesis testing for the presence of watermarks, a topic that is now well studied. In contrast, the finer grained task of localizing which segments of a text are watermarked is much less explored; existing approaches often lack scalability or guarantees robust to paraphrasing and post-editing. We bring a new perspective to this segmentation problem through the lens of epidemic change-points and, by exploiting this connection, propose WISER, a novel and computationally efficient watermark segmentation algorithm. We establish finite-sample error bounds and consistency for detecting multiple watermarked segments in a single text. Complementing these theoretical results, our extensive numerical experiments show that WISER outperforms state-of-the-art baseline methods, both in terms of computational speed as well as accuracy, on various benchmark datasets embedded with diverse watermarking schemes. Together, these theoretical and empirical results position WISER as an effective tool for watermark localization. Towards the end of the talk, we will briefly discuss spatial anomaly detection as another important application of the epidemic change-points in "space", and show how similar ideas lead to computationally fast algorithms with optimal detection-accuracy results for single and multiple patches. Based on joint works with Subhrajyoty Roy (Washington University at St. Louis), Sayar Karmakar (University of Florida), and George Michailidis (UCLA).