FL4SDN: A Fast-Convergent Federated Learning for Distributed and Heterogeneous SDN

The practical deployment of federated learning in distributed software-defined networks has to face many challenges, among which designing effective methods to handle heterogeneous environments with a convergence guarantee is of great concern. This paper proposes a novel algorithm called FL4SDN, using weighted averaging stochastic gradient descent updates for the federated optimization problem. In FL4SDN, the weight given to each controller is determined based on the relationship between the local and global gradients. Theoretically, it has been proven that faster convergence can be achieved if the loss function is convex and has a bounded first derivative. The effectiveness of FL4SDN is confirmed by experiments carried out on different network topologies. We demonstrate that such an algorithm can give a reduction of at least 24.4% in terms of communication rounds as compared to the state-of-the-art Gaia, FedAvg, Deep Gradient Compression, and Bulk-Synchronous Parallel. It also provides high prediction accuracy and low communication costs.