Fairness in Federated Learning

A minimax optimization framework that explicitly targets the worst-case performance across all client distributions. Instead of minimizing average loss, AFL minimizes the maximum loss, ensuring no single client or group is catastrophically left behind.

• Mechanism: Solves a minimax objective: min_w max_λ Σ λ_k * L_k(w)
• Key Benefit: Provides a theoretical guarantee on the upper bound of loss for the worst-performing client
• Trade-off: Often results in lower average accuracy compared to standard FedAvg
• Use Case: Critical applications like medical diagnostics where a model must not fail on a rare demographic

A flexible objective that tunes the degree of fairness by raising each client's loss to a power q during aggregation. A higher q forces the global model to distribute accuracy more uniformly across clients.

• Mechanism: Minimizes Σ (L_k^{q+1})/(q+1), dynamically re-weighting clients with higher loss
• q=0: Reduces to standard FedAvg, optimizing for average performance
• q→∞: Approaches AFL's minimax fairness, prioritizing the worst-off client
• Practicality: Allows a continuous dial to navigate the fairness-utility trade-off without full minimax optimization

A multi-task learning approach that balances global fairness with local personalization. Ditto adds a proximal term to each client's local objective, encouraging the local model to stay close to a globally fair model while adapting to local data nuances.

• Dual Objective: min_w L_k(w) + (λ/2) * ||w - w_global||²
• Key Insight: Pure global fairness can harm local accuracy for outlier clients; Ditto allows controlled deviation
• Hyperparameter λ: Controls the trade-off between global fairness adherence and local personalization
• Advantage: Avoids the 'one-size-fits-all' pitfall of purely global fairness methods

An approach that tackles fairness at the system level by optimizing the selection and weighting of clients during training rounds, rather than modifying the objective function.

• Mechanism: Prioritizes clients with historically higher loss or those from underrepresented groups during client selection
• Dual Fairness: Addresses both performance fairness (model accuracy) and representation fairness (participation in training)
• Implementation: Modifies the server's aggregation weights based on a fairness metric like the Gini coefficient of client accuracies
• Benefit: Can be layered on top of standard FedAvg without changing local training procedures

Extends adversarial debiasing to the federated setting by training a global model with a competing adversary that attempts to predict sensitive attributes from the model's representations.

• Architecture: A shared encoder with two heads—a task predictor and a sensitive attribute adversary
• Gradient Reversal: The adversary's gradient is reversed before updating the encoder, actively removing sensitive information from the latent representation
• Federated Challenge: Requires careful orchestration to ensure the adversary is trained on a balanced distribution of sensitive attributes across clients
• Outcome: Produces a representation that is maximally informative for the task but minimally informative for predicting protected attributes

A hierarchical strategy that first clusters clients with similar data distributions or fairness requirements, then learns a separate fair model for each cluster before distilling them into a globally equitable model.

• Process: 1) Cluster clients by distribution similarity, 2) Train cluster-specific fair models, 3) Ensemble or distill into a global model
• Assumption: Clients within a cluster have compatible fairness constraints, reducing intra-cluster conflict
• Benefit: Mitigates the problem of conflicting fairness definitions across clients with fundamentally different data distributions
• Application: Multi-national deployments where legal fairness definitions differ by jurisdiction