Research On Deep Learning Recognition Of Low Signal-To-noise Ratio And Few-shot Wireless Signals

With the explosive growth of wireless applications,massive amounts of data are interoperable between wireless devices,which puts forward higher requirements for identifying user’s identity to ensure information security.Traditional password-based authentication schemes require frequent key exchanges,which are difficult to adapt to dynamic wireless networks with large connections,and with the improvement of computing power,there is a risk of being compromised or leaked.As a cross-protocol endogenous security mechanism based on the physical layer,radio frequency(RF)fingerprint identification has become an effective solution for wireless device identification because of its uniqueness and the third party non-counterfeiting by using the hardware tolerance of the device.In a complex wireless network environment,low signal-to-noise ratio(SNR)and small sample scenarios are key difficulties that must be resolved in the application of wireless signal fingerprint recognition technology.The transmit power of wireless devices is low,and the SNR of the signal obtained at the receiver is low.In response to this situation,this paper proposes a RF fingerprint identification algorithm based on generative adversarial networks(RFGAN).The residual structure is used to better dig out the features from the interfered wireless signal with poor quality,and realize the wireless signal recognition under the condition of low SNR.The practical signal recognition results show that the average recognition rate of equipment in the indoor environment fluctuates around 99%.When the SNR is 5dB,the average recognition rate of 25 devices is 97.9%.In view of the situation that deep learning strongly relies on massive training data,and large batches of wireless signals are often difficult to obtain,this paper proposes few-shot wireless signal classification network based on deep metric learning(FSig-Net).The prior knowledge obtained by the feature extraction network of the FSig-Net can be used to constrain the complexity of the hypothesis space,achieve the purpose of few-shot recognition,and adaptively determine the sample similarity measurement method.When the number of training samples for a single device is 10,the average recognition accuracy of FSig-Net can reach 98.28%.In addition,this paper designs an individual identification software system for wireless devices,which can flexibly train various real-time acquired data to verify the performance of the above-mentioned identification algorithms.