Closed-Loop Fairness and Privacy Optimization for Federated Text Classification under Non-IID Data
Abstract
Abstract Federated text classification under non-IID data faces three coupled challenges: utility degradation, group fairness mismatch, and privacy leakage. To address these issues, this paper proposes FedMOJO, a closed-loop collaborative optimization framework for fair and privacy-preserving federated classification. On the client side, FedMOJO integrates low-rank representation learning, differentially private local training, and fairness--privacy collaborative optimization to reduce redundant sensitive information while preserving task-relevant semantics. On the server side, it performs low-rank structural alignment, global fairness-state fusion, state-driven control generation, and robust aggregation, thereby forming a cross-round closed loop between global coordination and local execution. Experiments on Twitter Sentiment, Bias in Bios, and Adult Income show that FedMOJO achieves a more favorable trade-off among utility, fairness, and privacy than representative federated baselines. Compared with DPFairFed, FedMOJO improves accuracy by up to 4.89 percentage points, while consistently reducing fairness disparity and privacy leakage and increasing MDL. These results indicate that closed-loop client--server coordination can effectively balance classification performance, group fairness, and privacy preservation without sharing raw data, providing a promising solution for privacy-sensitive federated learning scenarios.
Read the paper
Related studies
- Active Learning during Federated Learning for Object Detection
- Fairness-aware federated learning
- Decentralized Federated Learning
- Building a Personally Identifiable Information Recognizer in a Privacy Preserved Manner Using Automated Annotation and Federated Learning
- Preservation of Patient Level Privacy: Federated Classification and Calibration Models
- Efficient explainable federated learning
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.