Profiteer: A Proxy-Mediated Fairness Attack in Federated Learning for Tabular Data

Abstract

Federated learning (FL) enables collaborative model training across decentralized data sources while preserving privacy, yet its vulnerability to data-level fairness attacks remains underexplored. Prior attacks typically manipulate model updates, labels, or sensitive attributes, leaving unexplored a subtler vulnerability rooted in the local composition of tabular data. We present Profiteer, a proxy-mediated fairness attack for tabular FL that exploits statistical correlations among proxy features, sensitive attributes, and target labels. By oversampling local instances in which proxy features are associated with privileged-group membership and favorable outcomes, the attack skews the training distribution at compromised clients through a one-time data manipulation. Across benchmark datasets under both IID and non-IID settings, Profiteer consistently worsens group fairness while maintaining near-benign accuracy and remaining effective against evaluated defenses, with particularly strong impact under heterogeneous data distributions. Our results reveal a practical data-level vulnerability in tabular FL and underscore the need for defenses that account for proxy-mediated manipulation of local training data.