An Energy-Efficient Federated Learning Framework for Privacy-Preserving Intrusion Detection in IOT Networks
ID：78
View protection：Participant Only
Updated time：2026-07-22 16:09:43 Views：33
In-person
Abstract
Thousands of Internet of Things (IoT) devices are being deployed in various smart home, healthcare, industrial automation and smart city applications, creating a massive attack surface for cyber threats and the need for effective intrusion detection is greater than ever before. However, traditional centralized Intrusion Detection Systems (IDS) have a number of drawbacks when implemented in a heterogeneous IoT environment, such as privacy concerns, communication overhead, energy consumption, and scalability. While Federated Learning (FL) has become a very promising method for privacy-preserving model training, current FL-based IDS approaches typically ignore the limited resources of IoT devices and do not consider energy efficiency and communication optimisation aspects. Pursuing a robust cybersecurity infrastructure for IoT networks that is scalable, privacy-preserving and resource-efficient, the present research introduces an Energy-efficient federated learning-based intrusion detection system (EEFL-IDS) for privacy preserving IoT networks.
The ultimate goal of this research is to design an intrusion detection framework to improve detection accuracy, decrease energy usage, reduce communication overheads and maintain data privacy in distributed IoT environment. The proposed EEFL-IDS introduces Federated Learning, Lightweight Deep Neural Networks (LDNN), Energy-Aware Client Selection (EACS), model pruning, Knowledge Distillation, Top-K sparse parameter transmission and secure federated aggregation into a single framework, in order to achieve these goals. The framework allows for the formation of a global model for intrusion detection by collaborating a number of distributed IoT devices without losing raw network traffic data to the cloud. Moreover, an energy-based client participation mechanism is dynamically designed to select suitable devices for federated training, which helps to alleviate the computational load and prolong the operational lifetime of the devices.
To assess the performance of the proposed framework, the following benchmark IoT intrusion detection datasets were used: CICIoT2023, IoT-23, TON_IoT, and N-BaIoT. The experimental results reveal that the detection accuracy of EEFL-IDS is 98.31%, with a precision of 97.94%, recall of 97.68%, and an F1-score of 97.81%, which greatly surpass the conventional FedAvg IDS and other state-of-the-art federated intrusion detection systems. Moreover, the proposed model can decrease the energy consumption by 33.42%, the communication cost by 46.29%, the number of convergence rounds by 38.33% and the number of false alarms by 53.03% when compared with conventional federated learning-based IDS models. The improvements demonstrate the successful integration of energy-saving optimization with communication-efficient learning strategies within federated IDS.
The findings confirm the proposed EEFL-IDS's effectiveness in addressing the primary limitations of existing IDS frameworks, specifically in balancing intrusion detection accuracy, privacy protection, communication efficiency, and energy sustainability. This framework provides a scalable and practical cybersecurity solution for future IoT ecosystems. Upcoming research will focus on incorporating blockchain-based trust management, XAI, adaptive client clustering, and validation through real-world deployment to enhance the robustness and applicability of the proposed method.
Keywords
Federated Learning, Energy Efficiency, Privacy Preservation, Lightweight Deep Neural Network, Energy-Aware Client Selection, Cybersecurity
Post comments