Privacy-Aware Anomaly Detection in IoT Environments using FedGroup: A Group-Based Federated Learning Approach

Abstract

Concerns on the data security and privacy of smart home users have been growing popularity due to the rising usage of IoT devices. Many traditional machine learning techniques have been used to perform anomaly detections. However, these models need to send private IoT data to a central model for validation and training, raising security and efficiency issues. We propose a new Federated Learning (FL) method called FedGroup, which adopts the FedAvg method, but it updates the learning of the central model based on the learning changes brought by each group of IoT devices. Our experimental results showed that FedGroup achieved same or better anomaly detection accuracy compared to other federated and non-federated learning methods. Furthermore, we showed how ensemble learning may be used to connect many contributing models for superior average prediction performance. FedGroup also improve the detection of attack type detection and attack type detail detection. By comparing our new models with baseline models, our models performed better showing an accuracy of 99.64% accuracy with 0.02% FPR on attack type detection and 99.89% accuracy on attack type detail detection.