A Privacy-Preserving Intrusion Detection System for IoT Networks Using Federated Learning
TL;DR
The experimental results demonstrate that the proposed framework not only maintains competitive detection performance and data privacy but also reduces computational and memory requirements, supporting its suitability as an effective and lightweight IDS for resource-constrained IoT environments.
Abstract
With the increasing presence of IoT devices in the real world, this widespread presence leads to serious security challenges related to the privacy of these devices' data. Despite the important role of intrusion detection system (IDS) and its ability to identify malicious security activities in traditional centralized learning solutions that rely on collecting raw data from devices and sharing it directly to a central server, these solutions may raise concerns regarding data privacy and an increase in communication overhead. To address these challenges, this study proposes a privacy-preserving intrusion detection system using federated learning (FL) that enables distributed IoT devices to engage in collaborative learning, without share the raw data, only updates, with taking into account the preservation of data privacy, all existing IoT clients independently train the Multilayer Perceptron (MLP) models on their own data only, and then share only the models updates with the central server, the Federated Averaging (FedAvg) algorithm is used within the central server to aggregate the updates and create a global model. The proposed framework was evaluated using the NF-BoT-IoT dataset. The experimental results demonstrate that the proposed lightweight IDS achieves an accuracy of 83.21% and an F1-score of 81.78 %, with performance comparable to the centralized learning approach while preserving client data privacy. In addition, the proposed hybrid feature selection approach reduces the feature space from eight to five features, resulting in measurable computational benefits. In particular, leading to a 15.21% reduction in inference time, a 4.59% reduction in training time, and a 2.78% reduction in memory usage. These results demonstrate that the proposed framework not only maintains competitive detection performance and data privacy but also reduces computational and memory requirements, supporting its suitability as an effective and lightweight IDS for resource-constrained IoT environments.