MTL-PIE: A multi-task learning based drone pilot identification and operation evaluation scheme

Abstract

As one of the most promising industries, consumer-grade Unmanned Aerial Vehicles (UAVs), also known as drones, have changed our lives. Although significant progress in drones has been made, adversary impersonation attacks still pose severe risks to flying drones. In addition, authorized pilot miss-operations also has become a critical factor leading to drone flight accidents. To validate the pilot's legal status and remind the authorized pilot about their miss-operations, we propose a multi-task learning-based drone pilot identification and operation evaluation scheme named MTL-PIE. Specifically, we first present qualitative and quantitative guidelines to evaluate pilot operation proficiency. Then, we design a pilot identification module and an operation evaluation module to resist pilot impersonation attacks and assess pilot operation proficiency, respectively. Finally, we propose a soft-parameter sharing mechanism to transfer knowledge between two modules and a dynamic weight-adjusting algorithm to prevent domain-dominant problems. Numerical results show that MTL-PIE can verify pilot legal status with an accuracy of 95.36% (outperforming our previous work with a margin of 2%-3%) and act as assessors to evaluate pilot operation proficiency with an accuracy of 94.47%. Note that MTL-PIE needs only 35 ms to verify pilot legal status and assess pilot operation proficiency; it has great potential to reduce drone flight accidents.