On the Evidentiary Limits of Membership Inference for Copyright Auditing
This work addresses the reliability of copyright auditing tools for LLMs, showing that current methods are insufficient for legal evidence, which is an incremental but important finding for legal and AI safety communities.
The paper tackled the problem of using membership inference attacks (MIAs) as evidence in copyright disputes for large language models, and found that state-of-the-art MIAs degrade when models are fine-tuned on paraphrased data that preserves semantics, indicating they are not robust in adversarial settings.
As large language models (LLMs) are trained on increasingly opaque corpora, membership inference attacks (MIAs) have been proposed to audit whether copyrighted texts were used during training, despite growing concerns about their reliability under realistic conditions. We ask whether MIAs can serve as admissible evidence in adversarial copyright disputes where an accused model developer may obfuscate training data while preserving semantic content, and formalize this setting through a judge-prosecutor-accused communication protocol. To test robustness under this protocol, we introduce SAGE (Structure-Aware SAE-Guided Extraction), a paraphrasing framework guided by Sparse Autoencoders (SAEs) that rewrites training data to alter lexical structure while preserving semantic content and downstream utility. Our experiments show that state-of-the-art MIAs degrade when models are fine-tuned on SAGE-generated paraphrases, indicating that their signals are not robust to semantics-preserving transformations. While some leakage remains in certain fine-tuning regimes, these results suggest that MIAs are brittle in adversarial settings and insufficient, on their own, as a standalone mechanism for copyright auditing of LLMs.