FRECA: Fair, Robust and Efficient Client Contribution Evaluation in Federated Learning
Towards Fair, Robust and Efficient Client Contribution Evaluation in Federated Learning
AI summary
Researchers proposed FRECA, a method for evaluating client contributions in federated learning, and experiments show it can quantify each client's contribution accurately, efficiently and robustly. FRECA is based on the FedTruth framework to estimate the true update of the global model, balancing each client's contribution while filtering the influence of malicious clients; it can resist Byzantine attacks and integrates Byzantine-resistant aggregation algorithms. The method relies only on local model updates, requires no validation operations or datasets, and can be used for client selection and compensation.