Adaptive Lazily Aggregation Based on Error Accumulation

Federated Learning (FL) enables multiple clients to collaboratively train models without exposing their local data. FL is an effective approach to utilizing localized data while preserving clients’ data privacy, but it also brings significant communication overhead. To reduce communication overhead of FL, this paper proposes the Adaptive Lazily Aggregation based on Error Accumulation (EA-ALA) algorithm. It uses adaptive constraints to determine whether a client can skip a communication round with the server so as to diminish communication cost. It also adopts error accumulation to improve model accuracy. The experimental results on CIFAR10 and Fashion-MNIST datasets show that compared to vanilla FL, EA-ALA consumes only 52% and 61% of communication rounds to achieve higher model accuracy.