Fairness-Optimized Dynamic Aggregation (FODA)
Open Access DepositedA Systems Engineering Perspective on Equitable Federated Learning in Heterogeneous Environments
Federated Learning (FL) enables multiple organizations or devices to collaboratively train machine learning models without centralizing sensitive data. While FL addresses key privacy and governance constraints, it does not inherently guarantee equitable outcomes. In real deployments, federated clients often exhibit substantial heterogeneity. They differ in demographic composition, label prevalence, and data quality. Under these Non-IID conditions, standard aggregation methods can converge to operating points that systematically underperform for underrepresented or historically disadvantaged groups. This is an especially serious failure mode in high-stakes applications such as healthcare, finance, and public-sector decision support. This dissertation introduces Fairness-Optimized Dynamic Aggregation (FODA), a closed-loop aggregation framework that treats fairness as a continuously monitored system objective rather than a static post-hoc constraint. FODA combines (i) discrepancy-aware aggregation based on lightweight client-side distribution summaries (e.g., aggregated class-count vectors) and (ii) fairness-gap feedback derived from real-time evaluation signals. After an initial warm-up phase, the aggregation controller dynamically adjusts client weights across rounds to reduce observed group disparities while preserving predictive performance. The approach is designed to be compatible with privacy-constrained settings because it relies on aggregated statistics and model-level feedback rather than access to client raw data. FODA is evaluated in a simulated cross-silo FL environment using three tabular datasets (Adult Census Income, German Credit, and Heart Disease) under multiple data regimes, including IID baselines, protected-attribute skew, joint skew over protected attribute and label, and a feature-skew hidden bias stress test that blinds discrepancy signals while inducing unobserved client-side feature shifts. Across a comprehensive suite of performance metrics and group fairness measures (e.g., demographic parity, predictive parity, equal opportunity, average odds, equalized odds, and disparate impact), and against six representative baselines (FedAvg, FedProx, FedDisco, q-FFL, FairFed, and AgnosticFair), results show that FODA achieves competitive accuracy while improving fairness behavior and stability in the most stringent heterogeneity regimes. Collectively, these findings support a systems engineering perspective where maintaining fairness in heterogeneous federated environments benefits from dynamic control mechanisms that respond to observed disparities over time rather than relying solely on fixed aggregation rules.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.