Exploring features for membership inference in ASR model auditing

Francisco Teixeira, Karla Pizzi, Raphaël Olivier, Alberto Abad, Bhiksha Raj, Isabel Trancoso

Abstract: Highlights•Joint analysis of speaker- and sample-level threats in membership inference•In-depth study of feature sets for auditing privacy in ASR models•Gaussian and adversarial features to significantly boost MI effectiveness•Comprehensive evaluation of feature relevance for model auditing•Speaker/sample traits reveal key patterns in MI prediction accuracy