Membership Inference Attacks against Large Audio Language Models

The article presents a systematic evaluation of Membership Inference Attacks (MIA) against Large Audio Language Models (LALMs), demonstrating that common audio datasets show near-perfect train/test separability, indicating that MIAs may detect distribution shifts rather than model inference. The authors introduce a blind-baseline protocol to mitigate confounding effects and benchmark various MIA methods, revealing that memorization in LALMs is cross-modal, linked to a speaker's vocal identity and associated text. This research provides a standard for auditing LALMs, highlighting the importance of addressing distribution artifacts in MIA evaluations.