Can Watermarked LLMs be Identified by Users via Crafted Prompts?

Authors: Aiwei Liu, Sheng Guan, Yiming Liu, Leyi Pan, Yifei Zhang, Liancheng Fang, Lijie Wen, Philip Yu, Xuming Hu

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that almost all mainstream watermarking algorithms are easily identified with our well-designed prompts, while Water-Probe demonstrates a minimal false positive rate for non-watermarked LLMs. In our experiments, we demonstrate that the Water-Probe algorithm achieves high accuracy in detecting various types of watermarked LLMs. 4 EXPERIMENT ON WATERMARKED LLM IDENTIFICATION
Researcher Affiliation Academia 1 Tsinghua University 2 Beijing University of Posts and Telecommunications 3 The Chinese University of Hong Kong 4 University of Illinois at Chicago 5 Hongkong University of Science and Technology (Guangzhou)
Pseudocode Yes We provide the detailed steps of the Water-Probe algorithm in Algorithm 1 in the appendix.
Open Source Code Yes [Official]:https://github.com/THU-BPM/Watermarked_LLM_Identification
Open Datasets Yes For watermarked text detection, we used OPT-2.7B to generate texts on the C4 dataset (Raffel et al., 2020)
Dataset Splits No No explicit training/test/validation dataset splits are provided for the Water-Probe algorithm's evaluation or for the main LLM identification task. The C4 dataset is mentioned for generating texts in a separate watermarked text detection context, not for defining splits for the primary experimental setup.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running the experiments are provided in the paper.
Software Dependencies No The paper mentions using the 'Mark LLM (Pan et al., 2024) framework' but does not specify its version number or other software dependencies with their versions.
Experiment Setup Yes For all LLMs, the sampling temperature was set to 1, with the number of samples set to 104. ... We set ยต = 0.1 for our experiments.