Impacto das técnicas de equilíbrio de classes na privacidade: uma análise dos ataques de inferência de associação

Abstract

The use of machine learning models trained on unbalanced datasets with sensitive information has raised significant privacy concerns. One major issue is vulnerability to Membership Inference Attacks (MIAs), which aim to discover whether a particular data point was part of a model’s training set. In this context, it has been largely overlooked whether class imbalance handling methods — a prerequisite for training standard supervised classifiers on such data — exacerbate the disclosure of sensitive information. To address this gap, we conducted a series of empirical MIA evaluations on real-world imbalanced datasets to investigate this question. Our results show that techniques employed to balance the dataset for classification can significantly improve the success rate of MIA attacks. This has clear implications for researchers developing protective measures against such attacks, especially in scenarios with imbalanced datasets.