FedMDD: Multi-deliberation based calibration for federated long-tailed learning

Abstract

Federated learning is a decentralized framework enabling collaborative training of machine learning models across distributed data clients while ensuring privacy protection. Despite its advantages, traditional federated learning faces the global long-tailed imbalance, leading to poor performance by overemphasizing head classes and under-representing tail classes. While planned (pre-hoc) and post-hoc imbalance adjustments have been explored, post-hoc methods often require auxiliary data or suffer from overconfident decision boundaries, which limits their effectiveness. To address the overconfidence and out-of-distribution in existing solutions, we propose a multi-deliberation based post-hoc calibration method (FedMDD) tailored for the federated long-tailed problem. FedMDD calibrates the global decision boundary for balance. It incorporates a local–global feature contrast constraint to generate effective features and uses consistency across client models to deliberate a model-aware margin. This margin promotes a large relative distance between tail classes and the decision boundary, preserving privacy by leveraging model performance without requiring access to local class distributions. Extensive experiments demonstrate that FedMDD outperforms existing methods in balancing decision boundaries and enhancing privacy protection, achieving superior performance on long-tailed data distributions.