FedAttn-Credit: attention-augmented federated learning with adaptive differential privacy for rural inclusive finance...

Rural inclusive finance faces persistent challenges in credit assessment due to fragmented data ecosystems, heterogeneous borrower profiles, and stringent privacy constraints. This paper proposes FedAttn-Credit, a multi-party collaborative credit assessment framework that integrates horizontal federated learning with an attention-augmented scoring model and an adaptive differential privacy mechanism. The framework connects five categories of rural data holders-commercial banks, rural credit cooperatives, government platforms, agricultural e-commerce providers, and village-level microfinance institutions-through a central aggregation server that never accesses raw data. A multi-head self-attention module with group-wise importance gating enables the model to dynamically weight heterogeneous feature groups according to their contextual relevance for each borrower. An adaptive differential privacy strategy calibrates noise magnitude based on each participant's data volume and training-round gradient dynamics, providing stronger protection for small-sample participants without disproportionately degrading model utility. Experimental results on public microfinance records and a synthetic rural credit dataset show that FedAttn-Credit achieves an AUC of 0.8734, narrowing the gap to the centralized upper bound to 1.8 percentage points while outperforming all federated baselines. At equivalent accuracy, the adaptive privacy mechanism cuts cumulative privacy cost by roughly 47% relative to fixed-budget alternatives, a figure that we derive in "Privacy protection effectiveness and model performance tradeoff analysis" section from the Privacy Efficiency Ratio reported there (the two quantities express the same comparison from opposite directions). Ablation and robustness analyses confirm that the attention module, adaptive noise scheduling, and Shapley-based contribution evaluation provide compounding benefits under both standard and adversarial conditions.