A radio frequency fingerprinting general architecture, method and electronic device based on multi-modal large model knowledge distillation

Abstract

This invention relates to the field of wireless sensing and surveillance identification, specifically providing a general architecture and method for radio frequency fingerprint recognition based on knowledge distillation of a multimodal large model. This method is used for the extraction and identification of transmitter radio frequency fingerprint features, improving identification accuracy and reducing inference latency. First, wired mode baseband signals are acquired via fiber optic connection. Second, the multimodal signal data is preprocessed using data slicing, domain transformation, and RGB image generation. Then, the preprocessed wired and wireless mode baseband signals are input into a multimodal large model, and radio frequency fingerprint features are extracted through unsupervised learning. Next, knowledge distillation transfers the knowledge from the multimodal large model to a lightweight network model. Finally, the lightweight network model is fine-tuned for wireless scenarios and deployed to edge IoT devices for IoT device identification and authentication. This solution improves the accuracy, robustness, and model scalability of radio frequency fingerprint recognition.

Images