Federated Learning-Enabled Intelligent Routing Optimization for AI-Native SD-WAN Architectures
Contributors
Dr. Ashish Dixit
Dr. Vipin Sharma
Keywords
Proceeding
Track
General Track
License
Copyright (c) 2026 Sustainable Global Societies Initiative

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Abstract
Enterprises now depend on SD-WAN to stitch together branch offices, data centers, and cloud workloads, yet most commercial controllers still pick paths using fixed rules or a single machine-learning model that only works once every branch has uploaded its raw traffic logs to one place — an approach that strains available bandwidth, exposes sensitive flow data, and leaves the whole network dependent on one controller staying healthy. We propose an AI-native SD-WAN design in which each branch runs its own reinforcement-learning routing agent and the branches jointly refine a shared policy through federated learning. Path selection is cast as a multi-objective decision process that simultaneously targets low latency, low jitter, low loss, and fair link usage that aggregation rule and convergence argument needed to train this distributed policy under realistically uneven (non-IID) branch traffic. Testing the design in a discrete-event SD-WAN simulator against OSPF, ECMP, and a conventional centrally trained DRL controller, across topologies ranging from 4 to 128 sites, the federated scheme cuts mean end-to-end latency by as much as 34.6%, cuts packet loss by as much as 41.2%, and needs roughly 38% less per-round uplink traffic than plain FedAvg — while its convergence is both faster and steadier than FedAvg, FedProx, and SCAFFOLD.