Research map / Federated learning and privacy / Personalized and heterogeneous FL
FairTrade: Achieving Pareto-Optimal Trade-Offs between Balanced Accuracy and Fairness in Federated Learning
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.
Proposes FairTrade, a federated learning framework that uses client-side fairness-constrained optimization and server-side multi-objective Bayesian optimization (qEHVI) to find Pareto-optimal trade-offs between balanced accuracy and demographic parity/causal fairness.
Problem
Fairness-aware federated learning methods report accuracy, which is misleading under class imbalance; the paper targets Pareto-optimal trade-offs between balanced accuracy and group fairness in FL.
Approach
Server-side multi-objective Bayesian optimization (differentiable qEHVI over a GP surrogate) tunes the client fairness-constraint regularization weight and learning rate to maximize global balanced accuracy and minimize discrimination; the framework is metric-agnostic and works with DP and FACE.
Key results
- On Adult with R3C split, FairTrade reports BA 0.77 and DP 0.001, whereas baselines such as Agnostic-Fair, FCFL, FedFB and FairFed fall to BA about 0.50 with DP near 0.
- Across five datasets, FairTrade achieves BA of 0.69 to 0.82 with low DP (e.g., KDD BA 0.82, DP 0.002), while SoTA fair FL baselines show BA near 0.5.
- FairTrade remains effective under causal fairness (FACE) and attribute-based client splits, reaching BA around 0.68 to 0.82 with low FACE values.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FairTrade instead of FedAvg | Federated learning · Adult (R3C) | BA: 0.79 -> 0.77 | Hurt |
| Use FairTrade instead of FedAvg | Federated learning · Bank (R3C) | BA: 0.79 -> 0.78 | Hurt |
| Use FairTrade instead of FedAvg | Federated learning · Default (R3C) | BA: 0.69 -> 0.69 | No clear effect |
| Use FairTrade instead of FedAvg | Federated learning · Law (R3C) | BA: 0.72 -> 0.68 | Hurt |
| Use FairTrade instead of FedAvg | Federated learning · KDD (R3C) | BA: 0.59 -> 0.82 | Helped |
| Use FairTrade instead of FedAvg | Federated learning · Adult (R3C) | DP magnitude: 0.409 -> 0.001 | Helped |
| Use FairTrade instead of FedAvg | Federated learning · Bank (R3C) | DP magnitude: 0.076 -> 0.019 | Helped |
| Use FairTrade instead of FedAvg | Federated learning · Default (R3C) | DP magnitude: 0.053 -> -0.010 | Helped |
| Use FairTrade instead of FedAvg | Federated learning · Law (R3C) | DP magnitude: 0.037 -> -0.008 | Helped |
| Optimize for FACE fairness with FairTrade | FairTrade · Adult (R3C) and Adult (Attr3C) | FACE: -0.0094; 0.0035 | Mixed |
Surprising findings
- Fairness-aware FL baselines with high accuracy (e.g., 0.93 on KDD) can have balanced accuracy near 0.5, i.e., act as random classifiers, making their low DP scores misleading.
- Optimizing for FACE also yields low DP, suggesting DP and FACE are not conflicting fairness notions in these experiments.
Limitations
- Balanced accuracy slightly declines relative to FedAvg on some datasets.
- Evaluated only on binary classification with a single binary sensitive attribute.
- Privacy claim of zero leakage relies on secure aggregation rather than formal analysis.
Experimental setup
Datasets: Adult Bank Default Law KDD
Benchmarks: Adult Bank Default Law KDD
Metrics: balanced accuracy accuracy demographic parity FACE
Baselines: FedAvg FF-SMOTE Agnostic-Fair FCFL FedFB FairFed
Builds on: FedAvg fairness-constrained optimization (Agarwal et al. 2018) qEHVI Fair-SMOTE
Closest papers by meaning
- FedFACT: A Provable Framework for Controllable Group-Fairness Calibration in Federated Learning (NeurIPS 2025 · similarity 0.91)
- FedLF: Layer-Wise Fair Federated Learning (AAAI 2024 · similarity 0.90)
- Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity (CVPR 2024 · similarity 0.90)
- Demystifying Local & Global Fairness Trade-offs in Federated Learning Using Partial Information Decomposition (ICLR 2024 · similarity 0.90)
- You Get What You Give: Reciprocally Fair Federated Learning (ICML 2025 · similarity 0.89)
- LoGoFair: Post-Processing for Local and Global Fairness in Federated Learning (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.