Research On Signal-oriented Adversarial Defense Algorithm

In recent years,the signal modulation style recognition based on deep learning has a greater improvement in performance and accuracy than the signal recognition under the traditional approach.However,in electronic countermeasures application scenarios,failure to correctly analyze and process the enemy’s signals in the face of a counterattack can have incalculable consequences,while research on countermeasure defense is not yet mature in the signal field.Therefore,in order to study the adversarial attack defense method for signal modulation pattern identification,this paper works on the following aspects.First,an adversarial defense algorithm based on data reconstruction is proposed,which uses a stitching technique to reconstruct the signal data using patches for signal data size characteristics,and calculates the Euclidean distance to select the appropriate patch block to replace the original sample from the self-built patch dataset to do away with perturbations.The model recognition rate of the algorithm before and after facing different perturbations in different counterattacks,after the signal data reconstruction defense,the added perturbations are eliminated by the reconstruction,and the model recognition rate can still maintain a more stable value in the face of different reasonable perturbations.When facing different attacks before and after,the defense samples after signal data reconstruction can be well fitted with the initial original sample amplitude and constellation plots.The experiments verify that the scheme proposed in this paper has good defense capability and migration.Secondly,an adversarial defense algorithm based on improved activation function is proposed to analyze the characteristics of ReLU function and design a bounded C0 discontinuity function,which combines the different characteristics of two activation functions,K-WTA fuzzy and destroy the gradient and LReLU to solve neuron death to improve the deep learning model recognition accuracy,and splice them together.This ensures a better adversarial defense effect without losing model recognition accuracy.Finally,this paper designs and implements a signal adversarial defense system,which uses SpringBoot framework for backend development and React framework for front-end presentation.The main modules of the system include user management module,data reconstruction module,activation function modification module and data management module.In this paper,the functions of each module of the system are described in detail.Through the testing of the signal adversarial defense system,it is verified that the system can effectively resist the adversarial attacks and proves the usability and effectiveness of the system.