Optimization of Networking Strategies in Decentralized Federated Learning

Networking plays a central role in Decentralized Federated Learning (DFL), where peer-to-peer communication replaces centralized orchestration. However, current DFL networking designs face challenges in terms of robustness against failures, communication efficiency under constrained bandwidth, and unlinkability to preserve privacy against traffic analysis. Addressing these challenges requires networking-layer optimizations that go beyond standard message exchange.

This thesis aims to investigate and implement optimization strategies for DFL networking that enhance system-level performance and privacy. Potential research directions include resilient overlay design to handle dynamic joins and failures, efficient communication protocols that reduce redundancy and delays, and anonymization or obfuscation mechanisms (e.g., mixnets, onion routing, dummy traffic) to improve unlinkability.

reference:

• Sandrin Hunkeler, Linn Spitz: Peer-Based Mixing for Decentralized Federated Learning; Universität Zürich, Communication Systems Group, Department of Informatics, Zürich, Switzerland, September 2025, URL: link.
• Timothy-Till Näscher: Design and Implementation of a Lightweight Decentralized Federated Learning Platform; Universität Zürich, Communication Systems Group, Department of Informatics, Zürich, Switzerland, August 2025, URL: link.

None