Skip to content
InfoResearchPreprint

AudioTrace: File-Level Membership Detection for Audio Large Language Models

Published
Record updated
View JSON

Summary

Researchers study audio-file-level membership inference, which asks whether a given audio recording was used to train an audio LLM without knowing the training tasks or annotations. They introduce AudioTraceBench, a controlled benchmark of 1,285 recordings with fine-tuning membership labels, and find that existing text- and video-based methods adapted to audio perform close to random guessing across three audio LLMs. Their proposed AudioTrace method reaches an AUROC of 0.60, rising with greater fine-tuning exposure or adaptation capacity.