MineLoRa: Markov Transition Fields and Adaptive Neural Networks for LoRa Device Radio Frequency Fingerprint Identification

With the development of LPWAN technology, LoRa, one of the typical representatives, is widely used in various communication scenarios. At the same time, it is also faced with problems such as difficulty in identifying the terminal equipment and the easy mistake and confusion. The industry has widely valued LoRa RF fingerprint recognition technology, which has become a security focus that cannot be ignored. Most recent research on terminal recognition has been carried out by multi-class feature fusion. Still, it needs help with high computational complexity, significant noise and other external interference factors, and harrowing feature extraction. We propose a time-frequency feature analysis method based on GAF and MTF. The time-domain features are transformed into two-dimensional images using the Gram Angle field and Markov transfer field. The spectral graph has both time-domain and frequency-domain characteristics, and the feature probability distribution in the images is intuitively analyzed. Finally, we design a deep neural network model based on CNN improvement. It can accurately extract and classify the features contained in multi-time frequency. Our experiments show that a 97% recognition rate can be achieved in an indoor environment, and a 90% recognition accuracy can be achieved in an outdoor environment in time.