Data-Free Knowledge Distillation for Privacy-Preserving Efficient UAV Networks
Guyang Yu · 2022
UAVs have been widely used in various fields such as military, rescue, and agriculture, where a significant number of data sets involve military secrecy and data privacy. Deep neural networks play a crucial role in these applications, but they require tremendous computational resources. Model compression and acceleration algorithms have been proposed to tackle this problem, which has a strong assumption that the training samples for the original network are available. However, in real-world applications, the training dataset is often inaccessible due to privacy issue and transmission constraints. In the paper, we propose a novel privacy-preserving efficient UAV framework by compressing the deep neural network paradigm in a data-free manner. Our method leverages data-free teacher-student learning for UAV recognition model compression, which preserves the source data privacy while achieving efficient model inference at the test phase. Extensive experiments demonstrate that our approach attains satisfactory results on visual object recognition using UAV networks.