FedResAttNet: a federated deep residual attention network for cross-institutional loan default prediction

The prevalent data silos among financial institutions severely constrain cross-institutional collaborative modeling and limit the improvement of default prediction performance. Traditional federated learning methods exhibit significant deficiencies in handling complex feature relationships in high-dimensional financial data and addressing data heterogeneity across institutions. This study proposes a federated deep residual attention network (FedResAttNet), which integrates deep residual connections with multi-head attention mechanisms through an embedded integration architecture. This approach achieves unified optimization of feature extraction and cross-institutional collaboration. The method designs a data quality-aware intelligent aggregation strategy that combines institutional data quality assessment, distribution similarity analysis, and adaptive weight calculation. This strategy effectively addresses parameter fusion challenges in heterogeneous data environments. The framework integrates differential privacy mechanisms to ensure institutional data security, constructing a comprehensive privacy-preserving federated financial risk control solution. Federated experiments based on the Give Me Some Credit dataset with five institutions demonstrate that FedResAttNet achieves 0.9342 and 0.8891 in area under the curve (AUC) and F1-score metrics, respectively. These results represent improvements of 1.86 and 2.37 percentage points compared to the optimal baseline methods. The performance loss under highly heterogeneous environments is only 6.0%, significantly outperforming the 12.3% loss of traditional federated averaging (FedAvg) methods. Under differential privacy constraints with ε = 1.0, the algorithm maintains 84–89% performance levels, validating its practical applicability. Ablation experiments further confirm that residual connections, attention mechanisms, and quality aggregation contribute 4.19%, 2.53%, and 1.86% performance improvements, respectively. This research provides an effective technical pathway for secure collaborative modeling among financial institutions. It holds significant importance for promoting practical applications of federated learning in financial risk control domains.