Signature Filtering Boosts LLM Watermark Detection in Weak Signal Settings.

Statistical watermarks are crucial for organizations to attribute the outputs of large language models (LLMs), but current detection methods often struggle with weak watermark signals, repetitive texts, or when watermarks have been edited. Researchers have developed a new approach called signature filtering, a lightweight module applied during the detection phase, which enhances watermark detection without requiring any modifications to the watermark embedding or text generation processes. Signature filtering operates by learning a small set of "signature" tokens. The presence of these tokens tends to render watermark tests unreliable. Before detection, these identified signature tokens are removed from the text. The signatures themselves are determined by solving a mixed-integer linear program on a small training dataset, with constraints designed to maximize the true positive rate of detection. Empirical evaluations across four prominent watermark families (Kgw, Sweet, Unigram, Exp), four benchmark corpora (C4, MBPP, HumanEval, Code-Search-Net), and six different LLMs (Opt-1.3b, Opt-6.7b, Llama2-13b, Llama3.1-8b, Qwen2.5-14b, Phi-3-medium-14b) demonstrate significant improvements. Using 2- and 3-gram signatures, detection rates in weak-signal and low-entropy contexts increased from 8-31% without filtering to 78-99% with filtering, while maintaining controllable and often negligible false positives. In stress tests involving scrambled sentences and 25-50% token perturbations, 2-gram filters for Kgw-style watermarks largely preserved the detection gains observed in clean text, frequently matching or surpassing advanced watermark detectors like WinMax. This makes signature filtering a simple, scalable, and model-agnostic addition for strengthening provenance checks of LLM-generated text in various information processing workflows.