A Self-supervised UAV Detection Method Based on Channel State Information

Abstract

Due to the widespread applications and potential security risks, unmanned aerial vehicle (UAV) detection has received increasing attention in recent years. Among the various methods, one promising method is wireless sensing with channel state information (CSI). However, most CSI-based approaches are supervised and cannot fully utilize unlabeled data. To address this issue, this paper proposes a self-supervised learning method based on CSI. Utilizing a Transformer architecture, the method employs two decoders for prediction and reconstruction tasks to learn the features of the CSI. The pretrained model can be fine-tuned for UAV detection tasks with a small number of training samples to leverage unlabeled data. In experiments across multiple task scenarios, the pretrained model has achieved up to 100% accuracy when fine-tuned with the entire dataset, and approximately 80% accuracy when fine-tuned with only 1% of the data, which has surpassed those models without pretrained parameters. These results demonstrate that our pretraining method can enhance the accuracy of UAV detection while reducing dependency on labeled data.