RETRACTED ARTICLE: Blockchain managed federated learning for a secure IoT framework

Jiayong Chai, Jian Li, Muhua Wei, Chuangying ZhuEURASIP 2023 · DOI · OpenAlex

Note: This article has been officially retracted by the publisher. This summary is provided for informational purposes regarding its technical claims.

Introduction

The proliferation of the Internet of Things (IoT) has led to the generation of vast amounts of sensitive data, particularly in medical and industrial sectors. While Federated Learning (FL) offers a privacy-preserving training mechanism by keeping data on local devices, traditional centralized FL architectures present single points of failure. Integrating blockchain can decentralize this process, but it often introduces significant communication latency and struggles with data sparsity in smaller networks.

Problem Statement

Existing blockchain-based FL systems face three primary hurdles:

  1. Data Sparsity: Small or new IoT clusters often lack sufficient local data to train robust models.
  2. Latency: Global blockchain consensus mechanisms are slow, especially when communicating over long distances (WAN).
  3. Privacy Risks: Frequent model updates can potentially be exploited to infer private local data.

Proposed Methodology: Cluster-Based FL

The proposed framework introduces a hierarchical "cluster-based" architecture to mitigate these issues. Instead of a single global blockchain, the network is partitioned into smaller, localized clusters.

Hierarchical Communication

  • Intra-cluster Protocol: Devices within a Local Area Network (LAN) perform high-frequency training and reach consensus rapidly.
  • Inter-cluster Protocol: Only aggregated "cluster-level" updates are exchanged between different clusters over a Wide Area Network (WAN). This significantly reduces the volume of long-distance traffic.

Validation and Security

The system employs a modified Two-Phase Commit (2PC) protocol. Before merging updates from remote clusters, a local cluster validates the incoming gradients against its own data. This "prepare" and "merge/discard" cycle ensures that only beneficial updates are integrated, protecting the local model from "poisoned" or irrelevant data. Furthermore, exchanging aggregated gradients rather than individual device updates provides an additional layer of privacy.

Experimental Results

The framework was evaluated using a simulated cloud IoT environment with two networks (A and B) acting as clusters, with node counts varying between 10 and 30.

  • Model Performance: The cluster-based model achieved a peak accuracy of 72.6%, starting from a baseline of 11%. The corresponding loss was reduced from 3.6 to 0.8.
  • Convergence Efficiency: The proposed model reached convergence in approximately 170 rounds, outperforming standard blockchain-based FL, which required over 200 rounds.
  • Latency Reduction: Localizing the majority of communication within clusters significantly improved time-to-consensus compared to flat global architectures.

Conclusion and Limitations

The cluster-based FL framework demonstrates potential for improving the efficiency and accuracy of decentralized IoT learning. By pooling resources across clusters, it addresses data sparsity while maintaining low latency through hierarchical communication. However, the computational demands of blockchain protocols remain a challenge for resource-constrained sensors. Future work should explore integration with Edge Computing to further optimize power and memory usage.

Disclaimer: Due to the retraction of the original source, these findings should be interpreted with caution regarding their empirical reliability.

Summary updated

Abstract

Abstract In this work, we present a blockchain-based federated learning (FL) framework that aims achieving high system efficiency while simultaneously addressing issues relating to data sparsity and the disclosure of private information. It is more efficient to build a number of smaller clusters rather than one big cluster for multiple networks. Blockchain-based FL is carried out in each cluster, with the model changes being compiled at the end of the process. Following that, the accumulated updates are swapped across the clusters, which, in practise, improves the updates that are accessible for each cluster. When compared to the extensive interactions that take place in blockchain-based FL, cluster-based FL only sends a limited number of aggregated updates across a substantial distance. This is in contrast to the extensive interactions that take place in blockchain-based FL. In order to conduct an analysis of our system, we have implemented the prototypes of both cluster and blockchain-based FL models. The findings of the experiments show that cluster-based FL model raise the accuracy goes upto 72.6%, and goes down to 11%. The loss goes upto 3.6 and goes down to 0.8. In addition, cluster-based FL model has the potential to hasten the convergence of the model, provided that the same quantity of data is input into it. The reason for this is due to the fact that during a training cycle, cluster-based FL model combines the computational resources of many different clusters.
BibTeX
@inproceedings{chai2023retracted,
  author = {Jiayong Chai and Jian Li and Muhua Wei and Chuangying Zhu},
  title = {{RETRACTED ARTICLE: Blockchain managed federated learning for a secure IoT framework}},
  booktitle = {EURASIP},
  year = {2023},
  doi = {10.1186/s13638-023-02311-x},
}

Loading related content...

Figure