Federated Support Vector Machines for Intrusion Detection in Wireless Sensor Networks
Tharindu Niranjan, Rathgama Guruge (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-12-05
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025120411281
https://urn.fi/URN:NBN:fi:tuni-2025120411281
Tiivistelmä
Wireless Sensor Networks (WSNs) are implemented in critical environments progressively, yet their vulnerability to network layer attacks creates significant security risks. Traditional centralized Intrusion Detection Systems (IDS) are unsuitable for WSNs due to their high communication overhead, susceptibility to single points of failure, and privacy concerns arising from centralized data collection. To address these challenges, this thesis proposes a Federated Learning-based Support Vector Machine (FL-SVM) framework that enables distributed and privacy-preserving intrusion detection without requiring raw data sharing.
The proposed framework is evaluated using the WSN-DS benchmark dataset, which contains representative routing layer attacks such as Blackhole, Grayhole, Flooding, and TDMA attacks. To overcome the computational limitations of sensor nodes and enable non-linear decision boundaries in a communication efficient manner, the study employs kernel approximation techniques using feature mapping. These techniques allow a linear SVM to approximate the performance of a centralized non-linear classifier.
Experimental results demonstrate that the centralized SVM model achieves 98.5% accuracy, while the proposed FL-SVM model attains 97.5% global accuracy, reflecting only a marginal 1% performance drop despite preserving data locality across distributed nodes. The framework shows strong resilience under non-IID data conditions and achieves stable cross collaborator convergence. A key finding confirms that optimized feature mapping successfully balances accuracy and communication efficiency, offering a practical IDS solution for WSNs. Furthermore, the analysis highlights that secondary compression via sketching provided negligible functional benefit on this task. This study validates the feasibility of deploying robust, communication effcient, and privacy-preserving machine learning in resource-constrained WSN environments.
The proposed framework is evaluated using the WSN-DS benchmark dataset, which contains representative routing layer attacks such as Blackhole, Grayhole, Flooding, and TDMA attacks. To overcome the computational limitations of sensor nodes and enable non-linear decision boundaries in a communication efficient manner, the study employs kernel approximation techniques using feature mapping. These techniques allow a linear SVM to approximate the performance of a centralized non-linear classifier.
Experimental results demonstrate that the centralized SVM model achieves 98.5% accuracy, while the proposed FL-SVM model attains 97.5% global accuracy, reflecting only a marginal 1% performance drop despite preserving data locality across distributed nodes. The framework shows strong resilience under non-IID data conditions and achieves stable cross collaborator convergence. A key finding confirms that optimized feature mapping successfully balances accuracy and communication efficiency, offering a practical IDS solution for WSNs. Furthermore, the analysis highlights that secondary compression via sketching provided negligible functional benefit on this task. This study validates the feasibility of deploying robust, communication effcient, and privacy-preserving machine learning in resource-constrained WSN environments.