Towards Federated Learning with Attention Transfer to Mitigate System and Data Heterogeneity of Clients

Abstract

Federated learning is a method of training a global model on the private data of many devices. With a growing spectrum of devices, some slower than smartphones, such as IoT devices, and others faster, such as home data boxes, the standard Federated Learning (FL) method of distributing the same model to all clients is starting to break down-- slow clients inevitably become strugglers. We propose a FL approach that spores different size models, each matching the computational capacity of the client system. There is still a global model, but for the edge tasks, the server trains different size student models with attention transfer, each chosen for a target client. This allows clients to perform enough local updates and still meet the round cut-off time. Client models are used as the source of attention transfer after their local update, to refine the global model on the server. We evaluate our approach on non-IID data to find that attention transfer can be paired with training on metadata brought from the client side to boost the performance of the server model even on previously unseen classes. Our FL with attention transfer opens the opportunity for smaller devices to be included in the Federated Learning training rounds and to integrate even more extreme data distributions.