Abstract
Federated Learning (FL) often faces a conflict between model performance and fairness, particularly in datasets with significant class imbalance. Traditional fair FL methods frequently rely on standard accuracy, which can be misleading in imbalanced scenarios where a model can achieve high accuracy by simply predicting the majority class. We present FairTrade, a framework that prioritizes Balanced Accuracy (BA) and utilizes Multi-Objective Bayesian Optimization (MOBO) to identify Pareto-optimal solutions. By optimizing for both high BA and low discrimination scores, FairTrade achieves superior trade-offs compared to state-of-the-art baselines across multiple real-world datasets, ensuring robust performance even in non-IID environments.
Introduction
In Federated Learning, data is siloed across different clients, often leading to significant class imbalance. A major flaw in current "fair" FL research is the reliance on standard accuracy as a performance metric. In imbalanced datasets, a model can achieve high accuracy and "perfect" fairness simply by predicting the majority class for everyone—a useless classifier. The real test of utility is Balanced Accuracy (BA), the average of sensitivity and specificity. Existing fair FL methods often degrade BA to 0.5 (random chance) to achieve fairness. FairTrade addresses this by maximizing BA while minimizing discrimination simultaneously.
The FairTrade Framework
FairTrade operates as a two-tier optimization system designed to navigate the "fairness-performance" tug-of-war.
- Client-Side (Local Optimization): Each client trains a local model using fairness-constrained optimization, penalizing violations of fairness boundaries like Demographic Parity.
- Server-Side (Multi-Objective Bayesian Optimization): The server receives local model updates and performance metrics. It employs Multi-objective Bayesian Optimization (MOBO) with a q-Hypervolume Improvement (qEHVI) acquisition function to tune hyperparameters like the learning rate and fairness regularization parameter.
This approach allows the system to search for the Pareto frontier—the set of solutions where fairness cannot be improved without sacrificing accuracy. Furthermore, the framework is metric-agnostic, supporting both statistical fairness and complex causal fairness measures.
Experimental Evaluation
The authors evaluated FairTrade on five diverse real-world datasets: Bank, Default, Adult, KDD, and Law. To simulate realistic FL environments, they used both Random (R3C) and Attribute-based (Attr3C) data distribution strategies.
As shown in the results, FairTrade consistently outperformed six major baselines. Notably, on the KDD dataset, while baselines struggled to maintain utility when enforcing fairness, FairTrade achieved a Balanced Accuracy of 0.82 with a negligible discrimination score of 0.002.
Key Insights
- The Accuracy Trap: High accuracy in fair FL is often a mirage caused by class imbalance. By reporting Balanced Accuracy, FairTrade exposes that many state-of-the-art methods produce near-random classifiers in these settings.
- Efficiency and Privacy: By using Bayesian Optimization, FairTrade reaches optimal trade-offs within few communication rounds. It maintains zero privacy leakage by relying on secure multi-party aggregation of model weights and metrics.
- Causal Flexibility: The framework effectively handles causal fairness (e.g., Average Treatment Effect) using propensity score matching, which is typically computationally prohibitive in FL.
Conclusion and Future Work
FairTrade represents a significant advancement in balancing utility and equity in decentralized learning. While currently focused on binary classification, future work could extend the framework to multi-class problems and intersectional fairness. Additionally, addressing potential bottlenecks in server-side surrogate model maintenance and ensuring robustness against malicious clients remain important areas for exploration.
Summary updated