EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

ShieldDFL: A Blockchain-Based Federated Learning Framework With Dual Privacy Protection and Reputation-Driven Consensus

Yali Cai, Xuetao Du, Chen Zhang, Meilun Li

IEEE Access · 2025 · 13 · 103931-103943

DOI: 10.1109/access.2025.3576261 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2025. Verfasst von Yali Cai und 3 weiteren Autor:innen. Die Arbeit wurde bisher 5-mal zitiert (Crossref) und nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In the industrial Internet of Things (IIoT) scenarios, federated learning (FL) provides a privacy-preserving solution for utilizing industrial data. At the same time, blockchain integration enhances trustworthiness in federated learning training. However, existing blockchain-based FL frameworks still face several critical challenges: 1) Current consensus mechanisms lack effective filtering of malicious devices, allowing low-quality participants to interfere with global model training and compromise model robustness; 2) Existing privacy budget strategies are overly simplistic, making it difficult to balance privacy protection between statistical queries and gradient updates—strong privacy protection reduces model accuracy, while weak protection fails to defend against poisoning attacks. To address these challenges, this paper proposes a blockchain-based federated learning framework with dual privacy protection and reputation-driven consensus, called ShieldDFL. This approach employs a hybrid consensus mechanism driven by LSTM-based reputation scoring to dynamically evaluate both short-term and long-term device contributions, enabling the precise selection of high-quality devices. At the same time, it introduces an innovative dual privacy budget mechanism that applies differential privacy separately to statistical queries and gradient updates, ensuring robust privacy protection while maintaining high model performance. Experimental results on the MNIST and CIFAR-10 datasets show that the proposed method reduces the probability of malicious devices entering the consensus pool to 1.5%, lowers the success rates of SAR and BASR attacks to 5.8% and 2.1% respectively, while maintaining high model accuracy of 98.1% on MNIST and 87.6% on CIFAR-10. Overall, the proposed framework effectively breaks through the security and privacy bottlenecks of blockchain-based federated learning, providing an efficient and scalable solution for decentralized and trustworthy collab-oration in IIoT scenarios.

Zitieren

APA

Cai, Y., Du, X., Zhang, C., & Li, M. (2025). ShieldDFL: A Blockchain-Based Federated Learning Framework With Dual Privacy Protection and Reputation-Driven Consensus. IEEE Access, 13, 103931-103943. https://doi.org/10.1109/access.2025.3576261

Vancouver

Cai Y, Du X, Zhang C, Li M. ShieldDFL: A Blockchain-Based Federated Learning Framework With Dual Privacy Protection and Reputation-Driven Consensus. IEEE Access. 2025;13:103931-103943. doi:10.1109/access.2025.3576261

BibTeX
@article{Cai2025,
  author = {Cai, Yali and Du, Xuetao and Zhang, Chen and Li, Meilun},
  title = {ShieldDFL: A Blockchain-Based Federated Learning Framework With Dual Privacy Protection and Reputation-Driven Consensus},
  journal = {IEEE Access},
  year = {2025},
  volume = {13},
  pages = {103931--103943},
  doi = {10.1109/access.2025.3576261}
}
Kostenlos starten →