Research On Intelligent Specific Emitter Identification

Specific emitter identification aims to accurately identify the target emitter based on its own unique fingerprint features,which are generated by the hardware differences of its transmitter and parasitized on its transmission signal in the form of unintentional modulation.As a core subject in electronic reconnaissance,it can provide valuable information for our army before and during wartime,so that our army can understand the enemy's dynamics and intentions in time,and then achieve the goal of knowing each other and defeating the enemy.With the evolution of modern warfare from information form to intelligence form,it is a general trend that specific emitter identification should be developed in the direction of intelligence.Although deep learning provides an opportunity for the intelligent development of specific emitter identification,existing explorations have never taken advantage of deep learning.The crucial reason is that researchers have not applied deep learning according to characteristics of specific emitter identification.Specifically,the modulation information contained in the transmitted signal can be divided into intentional modulation information and unintentional modulation information.The intentional modulation information representing system features of the target emitter is harmful to the identification and occupies a large proportion,while the unintentional modulation information representing corresponding fingerprint features is beneficial to the identification but occupies a small proportion.How to mine the imperceptible but effective unintentional modulation information hidden in the intentional modulation information has always been the bottleneck problem of specific emitter identification.Therefore,only by arming deep learning with an ability to break the imbalanced proportion between the intentional and unintentional modulation information can we completely change the stagnant situation in the intelligent development of specific emitter identification.To address the imbalanced issue between the intentional and unintentional modulation information,we proposes four schemes and puts them into practice by designing various deep neural networks.These schemes include regional proposal mechanism,adversarial decomposition mechanism,regional proposal and adversarial decomposition joint mechanism,and bilinear pooling and adversarial decomposition joint mechanism.The regional proposal mechanism filters out much intentional modulation information by detecting distribution regions of the unintentional modulation information,while the adversarial decomposition mechanism completely eliminates the intentional modulation information by decomposing the intentional and unintentional modulation information.The regional proposal and adversarial decomposition joint mechanism realize the precise mining of the unintentional modulation information and the effective suppression of the intentional modulation information by detecting the distribution regions of the unintentional modulation information and decomposing the intentional and unintentional modulation information.The bilinear pooling and adversarial decomposition joint mechanism replaces the region proposal mechanism with the bilinear pooling transformation,and organically integrates the distribution regions detection of the unintentional modulation information and the decomposition of the intentional and unintentional modulation information.The deep neural networks designed according to these schemes have achieved remarkable performances in real data experiments,which fully proves that these schemes can effectively solve the imbalanced problem between the intentional and unintentional modulation information.Additionally,a feature visualization method is also used to investigate operation mechanisms of these schemes and compare their advantages and disadvantages in detail.Our work has created a new school for the intelligent development of specific emitter identification,and also will provide useful reference for most researchers in this field.