A Multi-model Fusion Approach for Hidden Camera Detection in Smart City Scenarios

With the rapid development of smart city technologies, hidden cameras embedded in IoT (Internet of Things) devices — such as socket outlets, elevator buttons, and smart home appliances—pose severe threats to personal privacy. Traditional hidden camera detection methods rely on single modalities (e.g., radio frequency, lens reflection), which suffer from poor robustness in complex smart city environments, including multi-device signal interference and high camouflage levels. To address this issue, this paper proposes a multi-modal fusion approach integrating three core techniques: YOLOv8 (You Only Look Once version 8) for object localization, InSPyReNet for saliency region extraction, and optical flow analysis for subtle motion detection. A machine learning fusion model (Light Gradient Boosting Machine + Support Vector Machine ensemble) combines outputs of the three modalities to generate final detection results. This study constructs a dedicated dataset of hidden cameras in typical smart city scenarios (e.g., hotels, elevators, smart homes), containing 5,000+ images that cover varying illumination and camouflage levels. Experimental results show that the proposed approach achieves an F1-score of 0.5405, Precision of 0.4950, and Recall of 0.5952 at a confidence threshold of 0.95, outperforming single-modal methods. This work provides a practical solution for privacy protection in smart city security systems.