FedCust: Offloading hyperparameter customization for federated learning

Federated Learning (FL) is a new machine learning paradigm that enables training models collaboratively across clients without sharing private data. In FL, data is non-uniformly distributed among clients (i.e., data heterogeneity) and cannot be redistributed nor monitored like in conventional machine learning due to privacy constraints. Such data heterogeneity and privacy requirements bring new challenges for learning hyperparameter optimization as the training dynamics change across clients even within the same training round and they are difficult to be measured due to privacy. The state-of-the-art in hyperparameter customization can greatly improve FL model accuracy but also incur significant computing overheads and power consumption on client devices, and slowdown the training process. To address the prohibitively expensive cost challenge, we explore the possibility of offloading hyperparameter customization to servers. We propose FedCust, a framework that offloads expensive hyperparameter customization cost from the client devices to the central server without violating privacy constraints. Our key discovery is that it is not necessary to do hyperparameter customization for every client, and clients with similar data heterogeneity can use the same hyperparametersto achieve good training performance. We propose heterogeneity measurement metrics for clustering clients into groups such that clients within the same group share hyperparameters. FedCust uses the proxy data from initial model design to emulate different heterogeneity groups and perform hyperparameter customization on the server side without accessing client data nor information. To make the hyperparameter customization scalable, FedCust further employs a Bayesian-strengthened tuner to significantly accelerates the hyperparameter customization speed. Extensive evaluation demonstrates that FedCust achieves up to 7/2/4/4/6% better accuracy than the widely adopted one-size-fits-all approach on popular FL benchmarks FEMNIST, Shakespeare, Cifar100, Cifar10, and Fashion-MNIST respectively, while being scalable and reducing computation, memory, and energy consumption on the client devices, without compromising privacy constraints.