ADASS: Anti-Drone Audio Surveillance Sentinel via Embedded Machine Learning

The proliferation of drones at a consumer level enables malicious users to equip them with recording devices to snoop sensitive information. Drones may also be used to carry dangerous objects, such as explosives. Protecting critical infrastructures from the security and safety threats imposed by drones is therefore of uttermost importance to prevent potentially hazardous attacks. Internet of Things (IoT) technology represents an enabler of drone detection solutions thanks to its deployability over large areas which might be difficult to monitor otherwise. However, state-of-the-art drone detection technology is based on deep learning, which requires intensive resources and cannot be easily deployed in embedded devices. In this paper, we propose Anti-Drone Audio Surveillance Sentinel (ADASS), the first noise-based drone detection solution that can be implemented in IoT devices. The overall model is composed of a preprocessing phase that emphasizes the frequency of noise emitted by flying drones and a successive classification via a compressed Convolutional Neural Network. To validate ADASS, we implement it on an embedded system and compare its performance to that obtained with a non-compressed model. Our results show that the detection accuracy of ADASS is the same as obtained with a full non-compressed model