Abstract
Federated Learning (FL) frequently encounters a fundamental conflict when trying to balance fairness and robustness. Both rare data (essential for fairness) and mislabeled data (detrimental to robustness) exhibit high training loss, causing standard algorithms to either over-index on noise or ignore minority groups. We introduce FedPCA (Performance-Capacity Analysis), a novel framework that utilizes a label-free feature dispersion metric to distinguish "hard-to-learn" rare data from "misleading" mislabeled data. By mapping clients into a two-dimensional Loss-Dispersion space, FedPCA achieves high fairness without being derailed by label noise.
Introduction
In Federated Learning, achieving Rawlsian Max-Min fairness ensures that the global model performs well for all clients, including those with minority or "rare" data distributions. Simultaneously, the model must remain robust against clients providing noisy or mislabeled data.
As shown in Figure 1, existing FL algorithms struggle because rare data and mislabeled data produce similar high-loss patterns. A fairness-oriented algorithm might prioritize a noisy client, thinking it is a struggling minority, while a robust algorithm might discard a legitimate minority group as noise. This "homogeneity in loss patterns" creates a zero-sum game where improving fairness often sacrifices robustness, and vice versa.
Methodology: Performance-Capacity Analysis
FedPCA breaks this deadlock by introducing a second dimension to client evaluation: the Feature Dispersion Score ($S$). While Loss measures performance, Dispersion measures the model's internal capacity to separate features in a dataset, independent of the provided labels.
The 2D Mapping Strategy
In each communication round, the server collects a (Loss, Dispersion) pair from each client. This allows for a clear categorization:
- Common Data: Low Loss, High Dispersion (Easy to learn and already mastered).
- Rare Data: High Loss, Low Dispersion (Legitimate but difficult data that requires more model attention for fairness).
- Mislabeled Data: High Loss, High Dispersion (Data that should be easy to learn based on feature separation, but shows high loss—a clear indicator of incorrect labels).
As visualized in Figure 3, these three groups form distinct clusters in the 2D space. FedPCA uses a 3-component Gaussian Mixture Model (GMM) to automatically identify these groups at the server level.
Local and Global Optimization
Once identified, FedPCA applies targeted strategies:
- Global Aggregation: The server calculates weights that prioritize clients with low dispersion scores (the "hard" rare data) while discounting those identified as mislabeled.
- Local Refinement: Mislabeled clients employ either FedPCA-D (dropping noisy data) or FedPCA-HS (High-Confidence Sampling), where they only train on samples the model is highly certain about.
Results and Discussion
FedPCA was evaluated on CIFAR-10 and two medical imaging datasets (RSNA ICH and ISIC 2019). The results demonstrate that FedPCA consistently outperforms state-of-the-art "Fair-only" and "Robust-only" methods.
On the RSNA ICH brain hemorrhage dataset, FedPCA achieved a worst-case accuracy of approximately 80.4%. In contrast, vanilla FedAvg reached only 74.4%, and specialized fairness methods like FedISM saw their performance drop to 72.8% when noise was introduced. This confirms that by accurately identifying the source of high loss, FedPCA makes fairness and robustness mutually reinforcing rather than competitive.
Conclusion and Future Work
FedPCA demonstrates that "capacity" is the missing link in evaluating client contributions in heterogeneous FL environments. By moving beyond loss-based metrics, we can robustly achieve fairness even in the presence of significant label noise.
Current limitations include a dependency on a short "warm-up" phase to establish reliable feature representations and the assumption of a three-group client structure. Future research will explore more complex noise types and the application of performance-capacity analysis to non-classification tasks such as regression and object detection.
Summary updated