Intelligent Aggregation Federation: A Learning-Based Metasystem for Secure and Robust Federated Learning in Finance

Federated learning (FL) is a pivotal paradigm for financial institutions to collaboratively train models while preserving data privacy, yet standard frameworks are hampered by a triad of challenges: vulnerability to poisoning attacks, lack of algorithmic fairness, and performance degradation under data heterogeneity. These limitations arise from static, rule-based aggregation mechanisms. This paper proposes Intelligent Aggregation Federation (IA-Fed), a learning-based meta-system that transforms the central aggregator into an intelligent agent via three synergistic mechanisms. First, a Gradient Embedding and Contribution Topology (GECT) mechanism achieves deep, structural perception of client updates for precise anomaly detection. Second, a Gradient Dynamics Prediction and Correction (GDPC) module leverages sequence modeling for temporal behavior understanding and proactive risk intervention. Third, a Reinforcement Learning-based Robust and Fair Aggregation Strategy (RFAS) learns a dynamic policy to explicitly balance accuracy, robustness, and fairness. Extensive evaluations on three financial datasets demonstrate that IA-Fed consistently and significantly outperforms state-of-the-art baselines. Notably, IA-Fed maintains resilience under 40% malicious client ratios and reduces the Equalized Odds Difference by over three-fold compared to the strongest baseline while achieving the highest accuracy. IA-Fed provides an effective and responsible framework for building secure and robust financial federated learning systems.