MAPLE-Fed: A Multi-center Adaptive Differential-Privacy Federated Learning Algorithm for Secure Modeling of Sensitive Data

Modern applications in healthcare, finance and cross-institutional research increasingly require building predictive models from sensitive data that reside across multiple centers. However, multi-center data often exhibit challenges such as data imbalance, non-IID distributions, and strict privacy regulations, which limit the effectiveness of standard federated learning methods. Direct data sharing is often infeasible due to privacy, regulatory and institutional constraints. In this work, we propose MAPLE-Fed, a novel multi-center federated learning framework that tightly couples adaptive differential privacy with heterogeneity-aware strategies and secure model exchange to enable high-utility, privacy-preserving modeling of sensitive distributed data. MAPLE-Fed introduces (1) an adaptive per-center privacy budgeting mechanism that dynamically allocates differential privacy (DP) noise according to each centers data utility, sensitivity and contribution, maximizing global model performance under a fixed privacy budget; (2) a unified heterogeneity-aware clipping and fairness calibration strategy that mitigates the adverse effect of non-IID distributions and small-sample centers; (3) a hybrid secure aggregation+cross-site representation distillation pipeline, where encrypted updates protect gradients while teacher-student distillation aligns latent representations across centers without exposing raw data or labels; and (4) an analytical privacy-utility trade-off analysis with practical scheduling rules for budget decay and aggregation frequency. We validate MAPLE-Fed on multi-center benchmarks and demonstrate consistent improvements in accuracy and fairness compared to baseline DP-FedAvg and naively privatized federated approaches, while satisfying rigorous DP guarantees. MAPLE-Fed thus provides a practical, theoretically grounded path for collaborative modeling with sensitive multi-center data.