FedSym: A Symbolic Framework for Communication-Efficient Federated Learning

Although Federated Learning (FL) allows for collaborative and communication efficient training of a global model without sharing raw data, current FL algo rithms still face challenges in low-bandwidth scenarios such as in space or disaster relief. This project introduces FedSym, a novel FL framework based on Sym bolic Regression (SR) that distils local models into a set of symbolic expressions for communication-efficient and interpretable FL. Various UCI datasets and the C-MAPSS Jet Engine Simulated datasets were used to evaluate the framework. FedSym methods achieved up to 99.4% inbound and 98.6% outbound communi cation reduction when compared to FedAvg on real-world datasets from the UCI repository, albeit at the cost of slightly worse performance. However, performance degrades on multimodal and heterogeneous data settings.