Feature Skew Control for In-Network Federated Learning

Applying Federated Learning (FL) to real-world network environments, such as telecommunication networks, presents significant challenges. In this paper, we investigate the intrinsic problem of non-IID data training across distributed innetwork FL clients. More specifically, we focus on the feature distribution discrepancy in the use of in-network FL for infrastructure monitoring. While previous works have made notable progress in the design of aggregation functions that compensate strong polarization in data distributions, limited attention was directed toward data load-balancing from sources to processing nodes, which leverages the role and characteristics of the data itself. Our goal is to address feature heterogeneity in in-network federated learning by piloting how data is forwarded in the network on the way to its consumers. To this end, we design and evaluate several scenarios for dynamic data load balancing between FL clients within the network topology, subject to latency constraints.