FedPCD
A fairness-oriented federated learning optimization framework that decouples the global optimization objective from client-level protection mechanisms through a priority constraint mechanism. It aims to improve the performance of disadvantaged clients while maintaining overall global accuracy.
Timeline (last 90 days)
The core design of FedPCD (Priority-Constrained Federated optimization) is to explicitly separate the global optimization objective from the client protection mechanism, retaining traditional federated aggregation as the primary objective while applying only the minimum correction necessary to satisfy client performance constraints.
FedPCD uses constrained optimization modeling to limit the correction to exactly what is needed to satisfy the constraints, thereby minimizing interference with the primary objective.
Across multiple heterogeneous federated learning benchmarks, FedPCD achieves the highest global accuracy under all configurations.
FedPCD leads on the fairness metric of worst-10% client accuracy in nearly all configurations.
FedPCD can reach performance comparable to or better than baselines in significantly fewer communication rounds.
All Facts (5)
The core design of FedPCD (Priority-Constrained Federated optimization) is to explicitly separate the global optimization objective from the client protection mechanism, retaining traditional federated aggregation as the primary objective while applying only the minimum correction necessary to satisfy client performance constraints.
50%UnverifiedFedPCD uses constrained optimization modeling to limit the correction to exactly what is needed to satisfy the constraints, thereby minimizing interference with the primary objective.
50%UnverifiedAcross multiple heterogeneous federated learning benchmarks, FedPCD achieves the highest global accuracy under all configurations.
50%UnverifiedFedPCD leads on the fairness metric of worst-10% client accuracy in nearly all configurations.
50%UnverifiedFedPCD can reach performance comparable to or better than baselines in significantly fewer communication rounds.
50%