Location
Havener Center, Miner Lounge / Wiese Atrium, 1:30pm-3:30pm
Start Date
4-2-2026 1:30 PM
End Date
4-2-2026 3:30 PM
Presentation Date
April 2, 2026; 1:30pm-3:30pm
Description
Cross-silo federated learning (FL) trains shared models across geographically distributed institutions without centralizing raw data, but its reliance on wide-area networks makes round completion time highly sensitive to heterogeneous link conditions, congestion, and straggler clients. Existing SDN-based FL frameworks address this through centralized traffic engineering, an approach that breaks down when silos span independent administrative domains where no single entity can maintain complete, timely network knowledge. We replace the centralized control plane with a fully distributed overlay in which each silo gateway operates as an equal peer, continuously probing local links, exchanging EWMA-smoothed metrics via bounded gossip, and computing least-cost routes independently on its evolving local graph estimate. Built on this shared network view, the FL server replaces uniform client sampling with a cost-weighted SoftMax selector that jointly minimizes predicted round latency and penalizes over-selected silos, balancing efficiency with long-run participation fairness. Trust scores derived from telemetry consistency further shape both routing and selection, providing graceful degradation under adversarial node behaviors and link failures without hard exclusion. Simulations across 100-silo topologies demonstrate up to 41% reduction in mean round completion time over uniform random selection.
Biography
Rabin is a graduate student in Computer Science at Missouri University of Science and Technology in Rolla, Missouri, working under the guidance of Dr. Md Arifuzzaman. Originally from Biratnagar, Nepal, he earned his bachelor's degree in computer engineering from IOE Purwanchal Campus, Dharan. His research interests lie at the intersection of Distributed Systems, Networking, and Distributed AI. Rabin is passionate about teaching and actively seeks opportunities to share knowledge within the academic community. Outside of his academic pursuits, he enjoys hiking, badminton, philosophy, birdwatching, and camping.
Meeting Name
2026 - Miners Solving for Tomorrow Research Conference
Department(s)
Computer Science
Document Type
Poster
Document Version
Final Version
File Type
event
Language(s)
English
Rights
© 2026 The Authors, All rights reserved
Included in
Distributed Control Plane for Cross-Silo Federated Learning
Havener Center, Miner Lounge / Wiese Atrium, 1:30pm-3:30pm
Cross-silo federated learning (FL) trains shared models across geographically distributed institutions without centralizing raw data, but its reliance on wide-area networks makes round completion time highly sensitive to heterogeneous link conditions, congestion, and straggler clients. Existing SDN-based FL frameworks address this through centralized traffic engineering, an approach that breaks down when silos span independent administrative domains where no single entity can maintain complete, timely network knowledge. We replace the centralized control plane with a fully distributed overlay in which each silo gateway operates as an equal peer, continuously probing local links, exchanging EWMA-smoothed metrics via bounded gossip, and computing least-cost routes independently on its evolving local graph estimate. Built on this shared network view, the FL server replaces uniform client sampling with a cost-weighted SoftMax selector that jointly minimizes predicted round latency and penalizes over-selected silos, balancing efficiency with long-run participation fairness. Trust scores derived from telemetry consistency further shape both routing and selection, providing graceful degradation under adversarial node behaviors and link failures without hard exclusion. Simulations across 100-silo topologies demonstrate up to 41% reduction in mean round completion time over uniform random selection.
Comments
Advisor: Md Arifuzzaman, marifuzzaman@mst.edu