FedNAS: Federated Deep Learning via Neural Architecture Search

Chaoyang He, Erum Mushtaq, Jie Ding, Salman Avestimehr

Abstract: Federated Learning (FL) is an effective learning framework used when data cannotbe centralized due to privacy, communication costs, and regulatory restrictions.While there have been many algorithmic advances in FL, significantly less effort hasbeen made on model development, and most works in FL employ predefined modelarchitectures discovered in the centralized environment. However, these predefinedarchitectures may not be the optimal choice for the FL setting since the user datadistribution at FL users is often non-identical and independent distribution (non-IID). This well-known challenge in FL has often been studied at the optimizationlayer. Instead, we advocate for a different (and complementary) approach. Wepropose Federated Neural Architecture Search (FedNAS) for automating the modeldesign process in FL. More specifically, FedNAS enables scattered workers tosearch for a better architecture in a collaborative fashion to achieve higher accuracy. Beyond automating and improving FL model design, FedNAS also provides anew paradigm for personalized FL via customizing not only the model weightsbut also the neural architecture of each user. As such, we also compare FedNASwith representative personalized FL methods, including perFedAvg (based on meta-learning), Ditto (bi-level optimization), and local fine-tuning. Our experiments ona non-IID dataset show that the architecture searched by FedNAS can outperformthe manually predefined architecture as well as existing personalized FL methods.To facilitate further research and real-world deployment, we also build a realisticdistributed training system for FedNAS, which will be publicly available andmaintained regularly.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2004.08546/code)