Federated Learning with Differential Privacy for Cross-Institutional Datasets: A Trade-off Analysis between Model Performance and Data Leakage

Collaborative machine learning across cross-institutional datasets, such as multi-hospital Electronic Health Records (EHR) and cross-bank financial fraud telemetry, is routinely constrained by stringent data protection mandates (e.g., HIPAA, GDPR) and competitive confidentiality concerns. While Federated Learning (FL) enables collaborative training without raw data centralization, standard FL architectures remain vulnerable to privacy-leaking attacks, including Membership Inference Attacks (MIA) and model inversion vector extraction. Integrating Differential Privacy (DP) into FL offers mathematically bounded privacy guarantees (ε, δ), but injects stochastic noise into local gradient updates, giving rise to a fundamental trade-off between model utility (AUC-ROC / F1-Score) and empirical privacy leakage. This paper presents an empirical trade-off analysis evaluating DP-FL across cross-institutional healthcare and financial datasets spanning 12 decentralized nodes over 100 global training rounds. We rigorously quantify how varying privacy budgets (ε ∈ [0.5, 10.0]) and gradient clipping thresholds (C ∈ [0.1, 2.0]) influence model convergence, utility degradation, and vulnerability against state-of-the-art MIA probes. Empirical results demonstrate that under a strict privacy budget (ε = 1.0), DP-FL reduces MIA vulnerability by 87.6% (capping attack accuracy near random guessing at 52.1%) while incurring an acceptable 4.8% utility drop in healthcare mortality prediction (AUC-ROC = 0.884 vs. 0.932 non-private FL). To mitigate the noise-utility penalty, we propose an Adaptive Noise-Decay DP-FL (AND-DPFL) Engine that dynamically recalibrates noise scale across training epochs, recovering 68.5% of lost utility while retaining strict privacy bounds (ε = 1.0, δ = 10⁻⁵).