Watermark Stealing in Large Language Models

Authors: Nikola Jovanović, Robin Staab, Martin Vechev

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We are the first to propose an automated WS algorithm and use it in the first comprehensive study of spoofing and scrubbing in realistic settings. We show that for under $50 an attacker can both spoof and scrub state-of-the-art schemes previously considered safe, with average success rate of over 80%.
Researcher Affiliation Academia Nikola Jovanovi c 1 Robin Staab 1 Martin Vechev 1 1Department of Computer Science, ETH Zurich.
Pseudocode No The paper describes its algorithm in prose and diagrams (e.g., Fig. 2) but does not include a formally labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes We make all our code and additional examples available at https://watermark-stealing.org.
Open Datasets Yes To query LMmo the attacker uses the C4 dataset s Real News Like subset (Raffel et al., 2020), also used in most prior work.
Dataset Splits No The paper references datasets used for evaluation and data collection for watermark stealing (e.g., C4 dataset, Dolly-CW), but does not explicitly provide training, validation, and test dataset splits for model training.
Hardware Specification No The paper mentions various language models (e.g., LLAMA-7B, MISTRAL-7B, GEMMA-7B) used in experiments, but does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running these experiments.
Software Dependencies No The paper does not provide specific version numbers for software dependencies such as programming languages, libraries, or frameworks used in the experiments.
Experiment Setup Yes We obtain n = 30,000 responses of token length 800;... We empirically set the clipping parameter c = 2 in all experiments;... In all experiments, for the watermarking schemes we use the default γ = 0.25 and δ = 4. We further generally use parameters ρatt = 1.6 and δatt = 7.5, tuning them on separate data (using only LMatt) when necessary.