A Two-Stage IoT Window Intrusion Detection System

Burglaries are still a global problem that affects victims not only financially but also psychologically. Therefore, more and more citizens rely on alarm systems to protect their property. Current commercial alarm systems come with the downside of requiring the user to arm or disarm the alarm system. Camera-based alarm systems come with privacy issues due to the monitoring of inhabitants. To overcome the downsides of current alarm systems, we propose using Inertial Measurement Units (IMUs) and a microcontroller to build a smart window burglary detection system for the Internet of Things (IoT). Our system uses a combination of Machine Learning and Deep Learning to automatically distinguish between attacks, such as window intrusions or intrusion attempts, and daily non-attack activities, such as opening, closing, or tilting a building window. Therefore, our system does not require user interaction. Since computational power and energy are very limited on the microcontroller (edge-computing), we propose to use a simple Decision Tree classifier on the microcontroller and offloading the more complex and more accurate Convolutional Neural Network classifier to a gateway. We performed real-world intrusion experiments on windows in collaboration with experts to build a novel training and evaluation dataset. The results show that our system is able to detect intrusions while successfully avoiding false alarms.