Researches On Individual Identification Technology Of Radio Emitter Based On Deep Learning

Individual identification of radio emitter refers to the process of extracting features from the received signal and determining the transmitter of the communication radiation source through a classifier.It is one of the core technologies in the fields of electronic countermeasures,wireless network security and spectrum management,and plays a huge role in both civilian and military fields.The mechanism of the radiation source fingerprint generation is due to the inevitable subtle "hardware defects" in the hardware circuit of the radio transmitter,which is manifested in the IQ data of the radiation source.Therefore,by analyzing IQ data of the radiation source,we can uniquely identify the wireless device.When using traditional machine learning methods to achieve signal classification,the classifier cannot automatically extract signal features,and all use expert features,which is difficult to adapt to the current complex electromagnetic environment.Since deep neural network can automatically extract and learn abstract high-dimensional features from data,it has great potential in the field of individual identification of radiation sources.Therefore,this thesis studies the application of deep learning network in individual identification of radiation sources.We build deep learning models with different structures to extract signal features and identifies corresponding radiation sources,which can effectively avoid various defects and difficulties faced when manually extracting features.Corresponding algorithms are designed for the following three application scenarios: individual identification of radiation sources for non-stationary distribution data,dynamic updating of models and incremental learning of radiation sources,and detection of unknown radiation sources.The thesis first studies the mainstream convolutional neural network and residual network models,focusing on improving the structure of Alex Net and Res Net.We built six deep learning models with different structures to extract the subtle features of the IQ signal of the radiation source,so as to realize the identification of radio frequency fingerprints.Based on open datasets,this thesis designs a data preprocessing method for slicing radiation source IQ sequences with a fixed-length sliding window.Under the condition that the IQ data of the radiation source satisfies the stable distribution,we compare the recognition accuracy and training time of different models with different data amounts,and the experimental results verify the availability of the improved model.Secondly,this thesis tests the radio frequency fingerprinting based on deep neural network in the scenario when the distribution of input data is non-stationary to time.Aiming at the problem of decreasing accuracy,this thesis designs a transfer learning algorithm based on model parameters.First,we need to use the data collected on the first day for pre-training,and then use the data collected on the second day to fine-tune the parameters of the last few fully connected layers of the pre-trained model.This method not only greatly improves the recognition accuracy,but also saves training time,which can effectively improve the robustness of the model on the task of RF fingerprinting.Finally,for the identification of emerging new radiation source data,this thesis designs a class incremental learning algorithm and a domain incremental learning algorithm by retraining the pre-training model,which can realize the dynamic update of the model.We use the improved CNN2 based on Alex Net as the pre-training model.Compared to the initial neural network algorithm,This incremental learning algorithm is more practical.For the problem that unknown radiation sources may appear in practical applications,this thesis designs an algorithm based on Open Max’s open set identification to realize the detection of unknown radiation sources.