Fair Federated Learning is a federated learning approach that ensures equitable model performance across all participating clients — preventing the scenario where the global model performs well on average but poorly for certain clients with minority data distributions.

Fairness Approaches

• AFL (Agnostic FL): Optimize the worst-case client loss — ensure no client is left behind.
• q-FFL: Assign higher weight to clients with higher loss — focus on underperforming clients.
• FedMGDA+: Multi-objective optimization — find Pareto-optimal solutions across all clients.
• Per-Client Thresholds: Set minimum performance thresholds for each client.

Why It Matters

• Equity: Without fairness constraints, majority clients dominate — minority clients get poor models.
• Manufacturing: A model that works for Tool A but not Tool B is unfair and operationally useless.
• Incentive: Clients won't participate in FL if the resulting model doesn't perform well for them.

Fair FL is no client left behind — ensuring the federated model performs well for every participant, not just on average.