Federated Fairness

Directly addresses the root cause of unfairness in federated networks: non-IID data distributions. Unlike centralized training, federated fairness algorithms must account for the fact that local client datasets represent fundamentally different populations.

• Client Drift: Mitigates the tendency of local models to diverge toward their own biased optima.
• Distribution Skew: Handles both label skew (different P(y)) and feature skew (different P(x)).
• Quantity Skew: Compensates for clients with vastly different volumes of local training samples.

A foundational technique that minimizes the worst-case empirical risk across all client distributions. Instead of optimizing for average global accuracy, Agnostic Federated Learning (AFL) optimizes for the client with the highest loss.

• Minimax Objective: Formally, the goal is to find weights w that minimize max_{client k} F_k(w).
• Guaranteed Floor: Provides a theoretical guarantee that no participating client will experience performance below a certain threshold.
• Trade-off: Often reduces average global accuracy to lift the performance of the worst-off distribution.

Frames federated fairness as a Pareto frontier problem, seeking a model where no client's performance can be improved without degrading another's. Algorithms like FedMGDA+ apply multi-gradient descent to find a fair Pareto-stationary point.

• Conflict Resolution: Dynamically resolves conflicting gradient directions between clients.
• Gradient Aggregation: Weights client updates to ensure the final aggregated gradient does not harm any specific group.
• No Explicit Trade-off Weights: Avoids the need for manually tuning fairness-accuracy hyperparameters.

Practical heuristics that adjust the influence of specific clients during aggregation. q-FedAvg dynamically scales client weights based on their local loss, giving higher priority to underperforming clients.

• Loss-Based Weighting: Clients with higher empirical loss receive proportionally larger weight in the global average.
• Fair Clustering: Groups clients with similar data distributions into cohorts and ensures balanced representation from each cluster.
• Tilted ERM: Uses exponential weighting of losses to focus training on high-loss data points across the federated population.

Distinguishes between fairness enforced locally on the device and fairness enforced globally by the aggregation server. Local fairness ensures a client's model is unbiased toward its own sub-populations, while global fairness ensures the aggregated model performs equitably across all clients.

• Demographic Parity: Ensures model predictions are independent of sensitive attributes across the entire federated cohort.
• Equalized Odds: Guarantees equal true positive and false positive rates for different groups, even when those groups are siloed across clients.
• Dual Constraint Optimization: Simultaneously applies fairness regularizers during local training and fairness-aware aggregation on the server.

Quantifies the performance gap between the best-served and worst-served client cohorts. Standard metrics include the variance of client accuracy and the coefficient of variation (CoV) of local test performance.

• Accuracy Parity: Measures the standard deviation of local model accuracy across all participating clients.
• Worst-Case Performance: Tracks the minimum accuracy among all clients as a direct fairness indicator.
• Jain's Fairness Index: Adapts a network resource allocation metric to measure the equity of model performance distribution.

The primary driver of fairness violations in federated settings. When client datasets are not independently and identically distributed, local optima diverge sharply. A model trained via standard FedAvg on label-skewed data—where one client holds 90% of Class A and another holds 90% of Class B—will systematically underperform on minority classes. Addressing non-IID distributions through robust optimization is a prerequisite for achieving federated fairness.

The phenomenon where local model updates diverge from the global optimum due to heterogeneous data distributions. In a fairness context, client drift causes the global model to be anchored to the majority distribution, erasing the signal from underrepresented clients. Techniques like FedProx add a proximal term to the local objective, explicitly penalizing deviation from the global model to mitigate this drift and preserve minority group performance.

A minimax optimization framework that explicitly targets fairness by minimizing the worst-case loss across any mixture of client distributions. Unlike standard empirical risk minimization, AFL optimizes for the maximum loss over a convex hull of client distributions, ensuring that no subpopulation suffers disproportionately poor performance. This is a direct mathematical formalization of the Rawlsian maximin principle in distributed learning.

A resource allocation-inspired approach that reweights the objective function to penalize clients with higher losses more aggressively. The hyperparameter q tunes the fairness-accuracy tradeoff: q=0 recovers standard FedAvg, while higher values of q force the optimization to focus on reducing the variance of performance across clients. q-FFL dynamically scales the step size based on local loss magnitude, ensuring high-loss clients exert proportionally greater influence on the global update.

A pragmatic alternative to enforcing uniform global fairness. Rather than forcing a single model to perform equitably across all distributions, personalization allows each client to fine-tune the global model on local data. Techniques like FedPer (federated learning with personalization layers) keep base layers shared while allowing client-specific classification heads, effectively decoupling representation learning from distribution-specific decision boundaries.

A fairness metric that measures the distance between learned feature representations for different client subgroups. In federated settings, models may learn to encode majority-group features with higher fidelity while compressing or distorting minority-group representations. Techniques like contrastive federated learning enforce uniformity in the representation space, ensuring that semantically similar inputs from different clients map to nearby embeddings regardless of the client's data volume.