Back to AI News
arxiv

Logit Laundering: Evading LLM Data-Use Auditing via Inference-Time Logit Reshaping

The paper examines watermark-based data-use auditing for LLMs from an adversarial angle, positing a non-compliant provider that trains on unauthorized data and tries to hide it from auditors. It proposes Logit Laundering, a framework that keeps model parameters unchanged and instead manipulates inference-time logits through a logit reshaping module with a plug-and-play statistical bias estimator to suppress watermark-induced signals. A confidence-aware scaling module adaptively calibrates suppression strength based on next-token prediction confidence to trade off evasion against model utility. The authors report evaluations showing evasion of several state-of-the-art auditing methods while preserving downstream task performance, and argue for more robust auditing approaches.

AI Brief

It gives AI practitioners a concrete adversarial threat model against watermark-based training-data auditing, useful for anyone relying on such audits for copyright or data-ownership compliance.

arXiv:2611.09446v1 Announce Type: cross Abstract: Data-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. However, systematic investigation from an adversarial perspective remains limited: non-compliant LLM providers would use unauthorized copyrighted datasets for training their own models to enhance model capabilities, while attempting to evade detection by auditors, thereby avoiding legal liability. To fill this research gap, we propose a novel evasion framework, termed Logit Laundering, where the non-compliant LLM provider manipulates model outputs at inference time to undermine detectable watermark signals. Specifically, we design a logit reshaping module to suppress watermark-induced statistical signals by integrating a statistical bias estimator as a plug-and-play component. This design keeps the model parameters unchanged and performs manipulation only in the logit space, thereby enabling auditing evasion while minimizing the impact on model's learned knowledge. To further balance evasion effectiveness and model utility, we introduce a confidence-aware scaling module to adaptively calibrate the suppression strength based on next-token prediction confidence. Extensive evaluations show that Logit Laundering can successfully evade existing various state-of-the-art auditing methods, significantly attenuating detectable signals, while preserving model utility on downstream tasks. Our work highlights the need for exploring more robust auditing methods to guard against evasion behaviors.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!