Topology Learning for Heterogeneous Decentralized Federated Learning over Unreliable D2D Networks

Zheshun Wu, Zenglin Xu, Dun Zeng, Junfan Li, Jie Liu

Introduction

Decentralized Federated Learning (DFL) is a promising paradigm for training machine learning models across mobile devices without a central server. By exchanging model updates only with local neighbors, DFL enhances privacy and reduces communication bottlenecks. However, real-world deployment faces significant challenges: data heterogeneity (non-IID data) and unreliable communication links. In device-to-device (D2D) networks, the use of User Datagram Protocol (UDP) often leads to packet loss and transmission outages, which can destabilize training.

System model of DFL over unreliable D2D links using UDP.

The ToLRDUL Framework

To address these issues, this research introduces ToLRDUL (Topology Learning considering the Representation Discrepancy and Unreliable Links). The core innovation lies in optimizing the network topology—the "who talks to whom"—to mitigate the combined effects of diverse local data and flaky connections.

The method involves three key components:

  1. Convergence Analysis: The authors derived a new convergence bound that explicitly accounts for "unreliable links-aware neighborhood discrepancy." This theoretical foundation shows how link variance acts as a penalty on training speed.
  2. Representation Modeling: Instead of sharing raw data, models represent local data distributions as Gaussian distributions. The discrepancy between neighbors is then measured using KL divergence, making the optimization mathematically tractable.
  3. Topology Optimization: Using the Frank-Wolfe algorithm, ToLRDUL periodically updates the mixing matrix to find a sparse, efficient topology. This topology prioritizes links that offer valuable diverse information while maintaining high reliability.

Experimental Evaluation

The researchers evaluated ToLRDUL using the CIFAR-10 dataset under two heterogeneity scenarios: Label Skew (Dirichlet distribution) and Feature Skew (Rotated images).

Performance comparison of ToLRDUL against baselines in terms of training loss and test accuracy for both label and feature skew.

Key Results

  • Superior Accuracy: On Dirichlet CIFAR-10, ToLRDUL achieved a test accuracy of 43.42%, outperforming STL-FW (41.10%) and random topologies (33.16%) after 300 rounds.
  • Enhanced Stability: As shown in the performance plots, ToLRDUL exhibits smoother convergence curves compared to baselines, which often suffer from "jitter" due to uncompensated packet loss.
  • Reduced Latency: By optimizing the communication graph, ToLRDUL significantly lowered transmission latency, achieving 65.73s in a 10-degree network compared to over 85s for competing methods.

Conclusion and Future Directions

ToLRDUL demonstrates that a carefully learned topology can act as a filter, balancing the need for diverse data with the constraints of unreliable wireless links. While the current work assumes a synchronous implementation and relatively static channels, future research could extend this framework to asynchronous DFL and highly dynamic mobile environments. Additionally, optimizing the local computational overhead for the Frank-Wolfe step remains an area for further refinement in low-power IoT contexts.

Summary updated

BibTeX
@article{wu2024topology,
  author = {Zheshun Wu and Zenglin Xu and Dun Zeng and Junfan Li and Jie Liu},
  title = {{Topology Learning for Heterogeneous Decentralized Federated Learning over Unreliable D2D Networks}},
  journal = {IEEE Trans. Vehicular Technology},
  year = {2024},
  url = {https://openreview.net/forum?id=s4pEhbFwSo},
}
Figure