Real-Time Federated Evolutionary Neural Architecture Search

Federated learning is a distributed machine approach to privacy preservation and two major technical challenges prevent wider application of federated learning. One that raises high demands on communication resources, since large number model parameters must be transmitted between the server clients. The other challenge training models such as deep neural networks in requires amount computational which may unrealistic for edge devices mobile phones. problem becomes worse when architecture search (NAS) carried out To address above challenges, we propose an evolutionary real-time NAS not only optimizes performance but also reduces local payload. During search, double-sampling technique introduced, each individual, randomly sampled submodel clients training. This way, effectively reduce costs required optimization, making proposed framework well suitable NAS.