Logit Laundering evades data-use auditing in large language models
Original title: Logit Laundering: Evading Data-Use Auditing in Large Language Models
Logit Laundering manipulates model outputs at inference time using logit reshaping to suppress watermark signals while preserving model utility on downstream tasks.
Logit Laundering is an evasion framework for data-use auditing in LLMs where non-compliant providers manipulate outputs to undermine watermark signals. It uses a logit reshaping module with statistical bias estimator and confidence-aware scaling. Evaluations show it evades various SOTA auditing methods while keeping model utility.
- logit reshaping suppresses watermark-induced signals
- confidence-aware scaling adapts suppression strength
- keeps parameters unchanged, manipulates only in logit space
- evades state-of-the-art auditing methods
- AI models
- LLM
From the original article
1 min read at arxiv.orgData-use auditing in Large Language Models (LLMs) aims to determine whether the given data has been used for training, thereby supporting data ownership safeguards. A dominant solution is data watermarking, which embeds traceable signals into training data and detects them in model outputs. Recent studies have proposed many watermarking-based auditing methods in LLMs.
By Ruihan Hu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.