PD-DRL: Towards privacy-preserving and energy-sustainable UAV crowdsensing

Abstract

Due to the high altitude advantage of unmanned aerial vehicles (UAVs), UAV crowdsensing has been extensively utilized in smart cities and harsh environments. However, UAVs have limited operational duration owing to energy constraints, dramatically diminishing their working efficiency. Moreover, their flight data is recorded and transmitted in unencrypted text, making it vulnerable to privacy breaches. We propose a privacy-preserving dual-model deep reinforcement learning approach (PD-DRL) to end it. It not only adaptively employs contextual knowledge to switch flight modes to enhance UAVs’ working efficiency but also safeguards the confidentiality of sensitive information during model training. PD-DRL consists of privacy-preserving deep reinforcement learning (P-DRL) and dual-model deep reinforcement learning (D-DRL). The former may integrate two distinct policies to switch between data collection and charging modes adaptively, hence optimizing the UAVs’ flight route. The latter can produce synthetic data to replace raw data during model training, thereby protecting the privacy of sensitive information. Ultimately, we conduct security discussions and comprehensive experiments to assess the effectiveness of PD-DRL. The discussion and experimental results demonstrate that PD-DRL surpasses other comparative algorithms, confirming its efficacy and safety.