An Explainable Multi-Task Learning Approach for RF-based UAV Surveillance Systems

Unmanned aerial vehicles or drones are ubiquitous among people, which can lead to technological, security, and community safety issues that must be addressed, monitored, and avoided. Intelligence services are always on the search for potential technology and intelligent systems that can identify drones. A potential drone surveillance system must be capable of detecting, localizing, identifying, recognizing the modes, and combating unauthorized drones. In this paper, we introduce a Multi-Task Learning (MTL) neural network for drone detection, identification, and drone mode detection using Radio Frequency (RF) signals. Due to the semantic abstraction of the drone RF signals, a single-task learning method can not fully meet the demands of the current anti-drone system. Moreover, executing each of the tasks, such as drone detection, type identification, and activity recognition, individually takes longer time, which is not applicable in a real-time drone surveillance system. Therefore, this paper proposes an MTL approach leveraging convolution layers to perform three tasks in parallel. A cross-entropy loss function used as the objective function optimization to improve the accuracy of the multiple tasks. The empirical results shows that the proposed MTL model achieve a better recognition accuracy compared to the existing solutions.