FEDERATED LEARNING-BASED ADAPTIVE ROUTING FOR PRIVACY-AWARE MULTI-DOMAIN NETWORK OPTIMIZATION

Abstract

Traditional routing algorithms depend on centralized data collection, which poses scalability and privacy challenges in large and distributed network environments. With the advent of Federated Learning (FL), routing models can now be trained collaboratively across multiple network domains without sharing raw traffic data, preserving privacy while maintaining global performance optimization. This paper presents a Federated Learning-based Routing (FLR) Framework that enables distributed routers or domains to collaboratively learn optimal routing policies. Each local model is trained using network telemetry such as delay, bandwidth, and packet loss, and only the learned parameters are shared with a global aggregator for model fusion. The aggregator applies Federated Averaging (FedAvg) to derive a global routing model that is redistributed to participating domains, enabling privacy-preserving and adaptive routing optimization. The proposed method ensures scalability and adaptability across heterogeneous domains. Simulation results show that the proposed FL-based routing achieves up to 26% lower average delay, 48% reduction in packet loss, and 7% higher throughput compared to centralized machine learning approaches, while significantly minimizing privacy leakage and communication overhead. The results validate that FLR can efficiently adapt to dynamic traffic patterns and heterogeneous domains, maintaining near-optimal routing performance without direct data exchange.