Research map / Federated learning and privacy / Personalized and heterogeneous FL

Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition

ICLR 2024 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Uses partial information decomposition to split global and local group-fairness disparity in federated learning into unique, redundant, and masked terms, derives fundamental trade-off limits, and proposes a convex AGLFOP optimization for the accuracy-fairness Pareto frontier.

Problem

Global and local group fairness in federated learning can disagree under data heterogeneity, and it is unclear when one implies the other or what accuracy-fairness trade-offs are information-theoretically achievable.

Approach

Defines global disparity as I(Z;Ŷ) and local disparity as I(Z;Ŷ|S), decomposes them via partial information decomposition into unique, redundant, and masked (synergistic) disparities, and formulates a convex AGLFOP over joint distributions with fixed marginals to bound achievable accuracy under global and local fairness constraints.

Key results

• Proves that local disparity can go to zero while global disparity stays positive whenever redundant disparity is positive, and that global fairness does not imply local fairness when masked disparity is positive.
• Shows AGLFOP is a convex optimization whose Pareto fronts on synthetic and Adult data reveal which fairness notion trades off against accuracy under each client split scenario.
• Experiments on Adult with FedAvg show unique disparity dominates when sensitive attributes are uniformly split across clients, while masked disparity dominates under high synergy splits.

Measured effects reported in the paper

Surprising findings

• Enforcing statistical parity at each client does not by itself remove redundant disparity; independence of predictions and clients is insufficient without zero masked disparity.
• Local fairness can be perfectly achieved while global disparity remains large, purely due to client-sensitive attribute correlation.

Limitations

• PID computation cost grows with number of clients or sensitive attributes
• Analysis restricted to binary sensitive attribute and binary label with independence fairness notion
• Experiments limited to synthetic data and Adult dataset

Experimental setup

Datasets: Adult synthetic 2-D Gaussian dataset

Benchmarks: Adult

Metrics: accuracy global disparity (mutual information) local disparity (conditional mutual information) statistical parity

Baselines: FedAvg with statistical parity regularizer

Builds on: partial information decomposition FedAvg Pinsker's inequality

Closest papers by meaning

• FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.92)
• FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning (AAAI 2024 · similarity 0.90)
• LoGoFair: Post-Processing for Local and Global Fairness in Federated Learning (AAAI 2025 · similarity 0.90)
• Does Egalitarian Fairness Lead to Instability? The Fairness Bounds in Stable Federated Learning Under Altruistic Behaviors (NeurIPS 2024 · similarity 0.89)
• Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity (CVPR 2024 · similarity 0.89)
• FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.88)