Research on federated learning against poisoning attacks for the Internet of Things
Year: 2025
Type: article
Abstract: In the Internet of Things （IoT） environment， federated learning is widely used for privacy protection and distributed computing， as it facilitates model training without the need for centralized data storage.However， the distributed nature of IoT devices and the diverse security requirements make federated learning systems vulnerable to data poisoning attacks， where attackers inject malicious gradients to disrupt the global model's training， thereby threatening system security.Although various defense strategies exist， ensuring robustness against such attacks while preserving privacy remains a challenge.To address this issue， this paper proposes a novel defense approach， named DPI， which detects and isolates malicious gradients using a lossless aggregation scheme that prevents the leakage of participants' actual gradients.Specifically， DPI first applies movable masks to each participant's gradients， then aggregates and performs dimensionality reduction using singular value decomposition.Subsequently， it employs a clustering algorithm to detect and discard poisoned gradients from the low-dimensional data.Experimental results show that DPI outperforms existing methods in detecting poisoned gradients， achieving a better balance between privacy protection and data security， and enhancing the robustness of federated learning systems in IoT environments.Moreover， DPI not only effectively mitigates poisoning attacks but also ensures the privacy of participants' gradients， thereby meeting the security requirements of IoT scenarios.
Language: Chinese
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162720849