Adaptive federated learning with non-IID data

Multiple clients can co-train high-performance global models using federated learning (FL) without revealing their personal information. The main obstacle to federated learning, however, is the considerable statistical variability across local data distributions of clients, which results in clients optimizing contradictory local models. We suggest a novel adaptive federated learning method with local drift decoupling and correction (AFEDDC) to tackle this basic problem. We experimentally demonstrate the efficiency, optimality and robustness of AFEDDC and show that AFEDDC outperforms existing algorithms and provides better convergence for FL.