Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging feature communication in federated learning for remote sensing image classification

Anh-Kiet Duong, Hoàng-Ân Lê, Minh-Tan Pham

TL;DR

This work addresses privacy and data-privacy concerns in federated learning for remote sensing image classification under non-IID data. It proposes feature-centric communication strategies, including sharing per-category average features and a hybrid model-and-feature approach, with Large Margin Cosine Loss (LMCL) to align representations. Key contributions include reducing data exchange to average features, introducing a regularized cross-client feature alignment, and presenting a retrieval-based deployment option that uses anchor features for prediction. Experiments on UC-Merced and AID demonstrate faster convergence and lower communication costs than standard FedAVG, with the best performance achieved by combining weight and feature communication alongside retrieval-based inference. The results offer a practical, privacy-preserving path for deploying FL in remote sensing scenarios with limited network bandwidth.

Abstract

In the realm of Federated Learning (FL) applied to remote sensing image classification, this study introduces and assesses several innovative communication strategies. Our exploration includes feature-centric communication, pseudo-weight amalgamation, and a combined method utilizing both weights and features. Experiments conducted on two public scene classification datasets unveil the effectiveness of these strategies, showcasing accelerated convergence, heightened privacy, and reduced network information exchange. This research provides valuable insights into the implications of feature-centric communication in FL, offering potential applications tailored for remote sensing scenarios.

Leveraging feature communication in federated learning for remote sensing image classification

TL;DR

This work addresses privacy and data-privacy concerns in federated learning for remote sensing image classification under non-IID data. It proposes feature-centric communication strategies, including sharing per-category average features and a hybrid model-and-feature approach, with Large Margin Cosine Loss (LMCL) to align representations. Key contributions include reducing data exchange to average features, introducing a regularized cross-client feature alignment, and presenting a retrieval-based deployment option that uses anchor features for prediction. Experiments on UC-Merced and AID demonstrate faster convergence and lower communication costs than standard FedAVG, with the best performance achieved by combining weight and feature communication alongside retrieval-based inference. The results offer a practical, privacy-preserving path for deploying FL in remote sensing scenarios with limited network bandwidth.

Abstract

In the realm of Federated Learning (FL) applied to remote sensing image classification, this study introduces and assesses several innovative communication strategies. Our exploration includes feature-centric communication, pseudo-weight amalgamation, and a combined method utilizing both weights and features. Experiments conducted on two public scene classification datasets unveil the effectiveness of these strategies, showcasing accelerated convergence, heightened privacy, and reduced network information exchange. This research provides valuable insights into the implications of feature-centric communication in FL, offering potential applications tailored for remote sensing scenarios.
Paper Structure (10 sections, 5 equations, 3 figures, 1 table, 4 algorithms)

This paper contains 10 sections, 5 equations, 3 figures, 1 table, 4 algorithms.

Table of Contents

  1. Introduction
  2. Method
  3. Revisiting FedAVG
  4. Feature-based FedAVG
  5. Model-and-features based FedAVG
  6. Retrieval-based classification
  7. Experiments
  8. Datasets and Setup
  9. Results
  10. Conclusion

Figures (3)

  • Figure 1: Visualization of Algorithm \ref{['alg:algo2']} with 2 categories (indicated by $+$ and $\times$) and 2 clients (in red and blue). The per-class mean feature vectors of each category ($\oplus$ and $\otimes$) is computed by each client and sent to the server which computes the average mean vectors (in green) before sending back to all the clients.
  • Figure 2: Comparative results when running Algorithms 1, 2, 3, 4, 5, and 6 on the UCM dataset.
  • Figure 3: Comparative results when running Algorithms 4, 5, 6 on the AID dataset.