Bias-Resilient Elephant Flow Detection in Distributed SDNs Through Federated Learning

This paper presents an innovative approach based on Federated Learning (FL) for detecting Elephant/Mouse Flows within distributed Software-Defined Networking (SDN) environments. In such networks, multiple domains are overseen by individual SDN controllers. Traditional centralized Machine Learning (ML) methods for detecting these flows involve training separate local models on data specific to each domain. This can lead to biased local models due to their dependence on context during data preparation. Consequently, these local models may struggle to adapt to different contexts and may not effectively handle the complexities of modern networks. To address these challenges, the proposed approach leverages the power of FL by fostering collaboration among controllers from various domains. This collaborative process aims to alleviate issues of local bias and, in turn, introduce greater adaptability and generalization to the detection process. Through experiments, we demonstrated the impact of data heterogeneity on the performance of local models and how the FL approach effectively mitigates these issues.