Open-set Identification Technology Of Communication Radiation Sources Based On Deep Learning

The specific emitter identification technology has a very wide range of application scenarios in the fields of electronic countermeasures and information security.Initially,people applied professional knowledge in the field of communication to manually extract the characteristics of communication radiation sources to identify individuals.With the increasingly complex wireless environment and the rapid development of artificial intelligence technology,the identification of communication radiation sources based on deep learning has entered people’s field of vision and achieved great success.However,most of these studies are based on a closed-set assumption that the categories of communication radiators in the test set must have appeared in the training set,and this assumption has been challenged in open environments.Therefore,this paper specifically studies the problem of open set identification of communication radiation sources,aiming to enable the neural network model to continuously improve the identification ability in the open environment and adapt to the requirements of the open environment.The specific research work consists of the following aspects:1.In order to make the neural network have the ability of open-set identification,that is,to have both the ability to classify known data and the ability to reject unknown data,an open-set identification method for radiation sources based on the Siamese network is proposed.This method ensures the integrity and distinguishability of extracted features through the limitations of classification loss,reconstruction loss and clustering loss,and then uses the Siamese network structure to judge the similarity of features,and then completes open set identification.By analyzing the Macro-F1 scores obtained in open set recognition,the effectiveness of all the work done in extracting features is verified.Compared with the classic open set recognition methods such as OpenMax and DOC,our method achieves better open set identification performance.2.In order to allow the neural network to further discover hidden classes in open-set data,a semi-supervised clustering method based on similarity transfer is proposed.The method consists of two parts,similarity transfer and semi-supervised clustering.According to the characteristics of communication radiation sources,a joint model is constructed to judge the similarity of unknown samples.The joint model is composed of classification network,similarity discrimination network and domain discrimination network.After the similarity matrix is obtained,it is regarded as the soft constraint and the distance between samples respectively,and the clustering is completed by two methods.Through the experimental analysis of similarity recognition accuracy,clustering accuracy and ARI index,it is proved that the method has good clustering performance.3.In order to effectively utilize the unknown samples clustered,a communication radiation source sample enhancement model based on CGAN is designed,which enhances the samples of unknown categories and continuously improves the open-set recognition ability of the model.The network consists of a generator and a discriminator.The two compete against each other to complete the training according to the given condition variables.Finally,samples of the specified category can be generated according to the condition variables.Finally,the real effectiveness of the generated samples is verified from multiple perspectives,and experiments are designed to verify the beneficial effect of sample enhancement on the improvement of open set identification performance.