Unveiling Privacy Risks in the Long Tail: Membership Inference in Class Skewness

Abstract

Real-world datasets often exhibit long-tailed distributions, raising important questions about how privacy risks evolve when machine learning (ML) models are applied to such data. In this work, we present a comprehensive analysis of membership inference attacks in long-tailed scenarios, revealing significant privacy vulnerabilities in tail data. We begin by examining standard ML models trained on long-tailed datasets and identify three key privacy risk effects: amplification, convergence, and polarization. Building on these insights, we extend our analysis to state-of-the-art long-tailed learning methods, such as foundation model-based approaches, offering new perspectives on how these models respond to membership inference attacks across head to tail classes. Finally, we investigate the privacy risks of ML models trained with differential privacy in long-tailed scenarios. Our findings corroborate that, even when ML models are designed to improve tail class performance to match head classes and are protected by differential privacy, tail class data remain particularly vulnerable to membership inference attacks.