Wireless Smart Home Security via Behavioral Profiling and Multimodal Representation Learning

The rapid proliferation of wireless-enabled smart home IoT devices over Zigbee, Wi-Fi, and MQTT networks has expanded the attack surface to include sophisticated spoofing attempts that exploit both behavioral and communication-layer weaknesses. This paper presents a behavior-driven spoof detection framework tailored for wireless smart homes, leveraging multimodal representation learning to profile device activities. Using the publicly available Smart Home Device Behavior Dataset, we extract domain-informed features—such as power efficiency, temporal usage patterns, and user-device interactions—and systematically inject synthetic spoofing behaviors to simulate adversarial conditions. Our pipeline applies Light Gradient Boosting Machine (LightGBM) classification with SMOTE oversampling, achieving 90\% accuracy, F1-score of 0.8893, and AUC of 0.9464. These results confirm that engineered behavioral features, when contextualized for wireless IoT environments, offer a lightweight and label-efficient approach for early spoof detection. Future extensions will integrate wireless-layer signatures (e.g., RSSI drift) and edge/federated deployment to address real-world wireless security constraints.