Research Article
Efficiently Detecting Anomalies in IoT: A Novel Multi-Task Federated Learning Method
@INPROCEEDINGS{10.1007/978-3-031-54531-3_6, author={Junfeng Hao and Juan Chen and Peng Chen and Yang Wang and Xianhua Niu and Lei Xu and Yunni Xia}, title={Efficiently Detecting Anomalies in IoT: A Novel Multi-Task Federated Learning Method}, proceedings={Collaborative Computing: Networking, Applications and Worksharing. 19th EAI International Conference, CollaborateCom 2023, Corfu Island, Greece, October 4-6, 2023, Proceedings, Part III}, proceedings_a={COLLABORATECOM PART 3}, year={2024}, month={2}, keywords={Internet of Things Multi-Task Federated Learning Anomaly Detection Feature Extractor Knowledge Transfer}, doi={10.1007/978-3-031-54531-3_6} }- Junfeng Hao
Juan Chen
Peng Chen
Yang Wang
Xianhua Niu
Lei Xu
Yunni Xia
Year: 2024
Efficiently Detecting Anomalies in IoT: A Novel Multi-Task Federated Learning Method
COLLABORATECOM PART 3
Springer
DOI: 10.1007/978-3-031-54531-3_6
Abstract
With the development of IoT technology, a significant amount of time series data is continuously generated, and anomaly detection of this data is crucial. However, time series data in IoT is dynamic and heterogeneous, and most centralized learning also suffers from security and privacy issues. To address these issues, we propose a multi-task anomaly detection approach based on federated learning (MTAD-FL) to address these problems. First, we propose a distributed framework based on Multi-Task Federated Learning (MT-FL), which aims to solve multiple tasks simultaneously while exploiting similarities and differences between tasks; second, to identify complex anomaly patterns and features in the IoT environment, we construct a Squeeze Excitation (SE) based and External Attention (EA) based Enhance Dual Network (SE-EA-EDN) feature extractor to monitor real-time data features from IoT systems efficiently; finally, we design a Local-Global Feature-based Parallel Knowledge Transfer (LGF-PKT) to parallelize the updating of weights of local and global features. To validate the effectiveness of our approach, we conducted comparative experiments on three publicly available datasets, SMD, SWaT, and SKAB, and MTAD-FL improved F1 by 11%, 67.8%, and 27.5%, respectively, over the other methods.
