Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

How Robust is Federated Learning to Communication Error? A Comparison Study Between Uplink and Downlink Channels

Linping Qu, Shenghui Song, Chi-Ying Tsui, Yuyi Mao

TL;DR

This paper investigates how robust federated learning is to digital communication errors, focusing on bit error rate effects in uplink and downlink channels. It develops a theoretical framework with standard assumptions (L-smooth nonconvex losses and bounded gradient variance) and derives explicit convergence bounds for both downlink and uplink BER via Theorem 1, revealing distinct BER dependencies. A key finding is that uplink BER impact is mitigated by aggregation across many clients and by the smaller range of model updates, while downlink BER impact scales with the larger parameter ranges. The theoretical results are validated through experiments on MNIST and Fashion-MNIST, demonstrating higher BER tolerance in the uplink and providing guidance for energy-efficient wireless FL design.

Abstract

Because of its privacy-preserving capability, federated learning (FL) has attracted significant attention from both academia and industry. However, when being implemented over wireless networks, it is not clear how much communication error can be tolerated by FL. This paper investigates the robustness of FL to the uplink and downlink communication error. Our theoretical analysis reveals that the robustness depends on two critical parameters, namely the number of clients and the numerical range of model parameters. It is also shown that the uplink communication in FL can tolerate a higher bit error rate (BER) than downlink communication, and this difference is quantified by a proposed formula. The findings and theoretical analyses are further validated by extensive experiments.

How Robust is Federated Learning to Communication Error? A Comparison Study Between Uplink and Downlink Channels

TL;DR

This paper investigates how robust federated learning is to digital communication errors, focusing on bit error rate effects in uplink and downlink channels. It develops a theoretical framework with standard assumptions (L-smooth nonconvex losses and bounded gradient variance) and derives explicit convergence bounds for both downlink and uplink BER via Theorem 1, revealing distinct BER dependencies. A key finding is that uplink BER impact is mitigated by aggregation across many clients and by the smaller range of model updates, while downlink BER impact scales with the larger parameter ranges. The theoretical results are validated through experiments on MNIST and Fashion-MNIST, demonstrating higher BER tolerance in the uplink and providing guidance for energy-efficient wireless FL design.

Abstract

Because of its privacy-preserving capability, federated learning (FL) has attracted significant attention from both academia and industry. However, when being implemented over wireless networks, it is not clear how much communication error can be tolerated by FL. This paper investigates the robustness of FL to the uplink and downlink communication error. Our theoretical analysis reveals that the robustness depends on two critical parameters, namely the number of clients and the numerical range of model parameters. It is also shown that the uplink communication in FL can tolerate a higher bit error rate (BER) than downlink communication, and this difference is quantified by a proposed formula. The findings and theoretical analyses are further validated by extensive experiments.
Paper Structure (14 sections, 23 equations, 10 figures)

This paper contains 14 sections, 23 equations, 10 figures.

Table of Contents

  1. Introduction
  2. System Model
  3. Federated Learning
  4. Federated Learning with Communication Error
  5. BER Tolerance: Uplink vs. downlink
  6. Intuitive Analysis
  7. Theoretical Analysis
  8. Experiments and Discussions
  9. Experiment Setup
  10. Results and Analysis
  11. Conclusion
  12. Acknowledgment
  13. Proof of Theorem 1
  14. Proof of Lemmas

Figures (10)

  • Figure 1: FL over noisy wireless channels.
  • Figure 2: The uplink error can be averaged in the server.
  • Figure 3: How the learning performance with downlink or uplink bit errors is affected by the number of clients.
  • Figure 4: The range of model parameters for MNIST experiment.
  • Figure 5: MNIST experiment: downlink tolerates BER of $10^{-4}$ while uplink can tolerate $10^{-1}$.
  • ...and 5 more figures