Asynchronous Federated Learning for Elephant Flow Detection in Software Defined Networking Systems

Abstract

This paper introduces an Asynchronous Federated Learning (AFL) approach to train an elephant flow model over Software Defined Networking (SDN) systems with distributed controllers. The AFL addresses the issues of data privacy and communication overhead in collecting network statistics over large-scaled SDN systems. It allows each local controller to train a local model based on its local statistics using Decision Tree and upload the local model to the root in an asynchronous manner, so that the root controller can aggregate each local model into a global model once a local model is received to improve its time efficiency. The AFL proposes to weight the performance of each local model to form the global model. The evaluation based on 5 real packet traces demonstrates the accuracy of the AFL is better than any local models and two classical federated learning approaches.