EUVIMEDEUVIMED entdecken
Crossref

FedPoison: A Dual-Strategy Poisoning Attack on Multi-Agent Federated Learning With Local Differential Privacy

Miao Du, Yinqiu Liu, Guangyuan Liu, Zhiqiang Kou, Zhu Han

IEEE Transactions on Cognitive Communications and Networking · 2026 · 12 · 9052-9065

DOI: 10.1109/tccn.2026.3706056 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Cognitive Communications and Networking, 2026. Verfasst von Miao Du und 4 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Gefördert von: Nanjing University of Posts and Telecommunications University-level Scientific Research Fund Project, Japan Science and Technology Agency (JST) Adopting Sustainable Partnerships for Innovative Research Ecosystem, University of Houston.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Federated learning (FL) is increasingly adopted in multi-agent healthcare systems where agents collaboratively train models without sharing sensitive data. While integrating FL with local differential privacy (LDP) offers robust protection against many privacy concerns, these models are still vulnerable to advanced adversarial attacks. A significant challenge is that even with LDP safeguards, FL models can be compromised by malicious attackers who inject poisoned data. In this paper, we introduce FedPoison, an advanced poisoning attack scheme designed to target both the mean and variance of datasets. Our method employs a dual strategy, integrating both input and output poisoning to effectively disrupt the FL process. Specifically, attackers can estimate the mean and variance of training data under LDP protection and inject poisoned data to carry out input poisoning attacks (IPA). The manipulated data in IPA is trained alongside the FL model, ultimately skewing the FL model’s output. Additionally, the attackers can bypass the LDP perturbation mechanism to launch output poisoning attacks (OPA) by uploading carefully altered mean and variance values, thus leading to biases in the aggregation of local model updates. Through the rigorous mathematical analyses, we demonstrate that the proposed scheme can successfully execute poisoning attacks on LDP-protected FL models. Extensive experiments conducted on four different datasets using PyTorch validate the effectiveness of our scheme, revealing that it significantly undermines the performance of FL models, even when LDP mechanisms are employed, highlighting its considerable threat to multi-agent collaborative decision-making systems in healthcare.

Förderung

  • Nanjing University of Posts and Telecommunications University-level Scientific Research Fund Project
  • Japan Science and Technology Agency (JST) Adopting Sustainable Partnerships for Innovative Research Ecosystem
  • University of Houston

Zitieren

APA

Du, M., Liu, Y., Liu, G., Kou, Z., & Han, Z. (2026). FedPoison: A Dual-Strategy Poisoning Attack on Multi-Agent Federated Learning With Local Differential Privacy. IEEE Transactions on Cognitive Communications and Networking, 12, 9052-9065. https://doi.org/10.1109/tccn.2026.3706056

Vancouver

Du M, Liu Y, Liu G, Kou Z, Han Z. FedPoison: A Dual-Strategy Poisoning Attack on Multi-Agent Federated Learning With Local Differential Privacy. IEEE Transactions on Cognitive Communications and Networking. 2026;12:9052-9065. doi:10.1109/tccn.2026.3706056

BibTeX
@article{Du2026,
  author = {Du, Miao and Liu, Yinqiu and Liu, Guangyuan and Kou, Zhiqiang and Han, Zhu},
  title = {FedPoison: A Dual-Strategy Poisoning Attack on Multi-Agent Federated Learning With Local Differential Privacy},
  journal = {IEEE Transactions on Cognitive Communications and Networking},
  year = {2026},
  volume = {12},
  pages = {9052--9065},
  doi = {10.1109/tccn.2026.3706056}
}
Kostenlos starten →