Abstract:
The rapid proliferation of Internet of Things (IoT) devices in smart home environments has amplified significant cybersecurity challenges due to their outdated firmware, limited computational resources, and weak protection mechanisms. Traditional centralized anomaly detection approaches are ineffective in these environments owing to single points of failure, privacy risks, and limited scalability. To address these challenges, this study proposes a novel federated learning framework that integrates a lightweight hybrid architecture combining 1D Convolutional Neural Networks (1D-CNN) with shallow autoencoders to support both supervised and unsupervised anomaly detection. The supervised component enables robust binary classification of known attack patterns, while the unsupervised reconstruction mechanism enhances detection of previously unseen anomalies, thus balancing accuracy with adaptability. Distinct from prior approaches, the proposed framework employs dynamic quantization within the federated learning process, enabling substantial reductions in resource consumption without compromising detection performance. Experiments conducted on the TON-IoT dataset, with SMOTE-based class balancing and PCA-driven dimensionality reduction, under a realistic non-IID distribution across three federated clients, demonstrate that the proposed model achieves 97.39% test accuracy. This model is designed for deployment on resource-constrained devices, employing post-training dynamic range quantization that reduces the model size by 81.5% (42.60 KB to 7.87 KB) without compromising accuracy. The experiment performed using Raspberry Pi 4 confirmed sub-millisecond inference and real-time responsiveness, validating the practical viability of the framework for efficient anomaly detection in smart IoT systems.
