FedNAS: A Distributed Neural Architecture Search
Tamai Ramírez-Gordillo, Higinio Mora, Francisco A. Pujol, Antonio Maciá-Lillo · 2024
In this paper, we present FedNAS, a distributed neural architecture search framework that integrates Federated Learning (FL) with Efficient Neural Architecture Search (ENAS) algorithm. This novel approach leverages the privacy-preserving capabilities of FL and the automated design optimization of NAS to collaboratively discover high-performance neural network architectures without sharing raw data. We propose a method wherein multiple clients independently execute the ENAS algorithm using local data and share their learned parameters with a central server. The server aggregates these parameters to create a global weight dictionary, facilitating knowledge transfer between clients and improving the overall model iteratively. We expect to demonstrate the potential of FedNAS by outlining its application to the CIFAR-10 dataset and discuss future work aimed at empirically validating the framework and comparing its performance against state-of-the-art architectures. Additionally, Blockchain technology will be integrated to enhance privacy and decentralization.