A Deep Learning Intrusion Detection Method With Resisting Adversarial Attacks For Industrial Wireless Sensor Networks

With the rapid development of Wireless communication,Industrial Wireless Sensor networks(IWSN)is a very important basic technology in the Industrial sector,Industrial core infrastructure of information sensing and transmitted by Wireless communication,real-time monitoring of the Industrial environment.However,a large number of industrial equipment,wireless communication industrial network,link control industrial system make industrial wireless sensor networks vulnerable to a variety of network attacks,leading to the intentional destruction or loss of important information and unable to timely transmission,which may cause irreparable industrial losses.To solve the problem of IWSN security,the need to guard against network attacks,Intrusion Detection system based on Deep Learning(DL-IDSs)stand out in the numerous safety protection means.DL-IDSs for network protection has brought great convenience,but in the face of the fight against security threats,cannot satisfy IWSN to safe protection measure on the accuracy and real-time requirements.Therefore,this article is based on the depth of the most popular learning model,from two aspects of against the attack and defense,is designed to resist against attack deep learning intrusion detection model,to ensure the safety of IWSN.The main research content is as follows:(1)Research under the industrial wireless sensor networks based on WGAN-GP method to generate counter sample integration,using the generated against network solution against training can only be a single attack and defense needs a large number of training samples of faults,improve the robustness of the neural network.Using real samples of natural gas transmission pipeline as against the attack of the original sample,using the Gradient punishment Wasserstein type generated against Network(WGAN-GP)several common fight for integrated training samples,with abundant sample set;Using multilayer perceptron Softmax classifier attack detection model is established,the validation method for generating WGAN-GP confrontation training effect,the model will be able to a variety of counter samples for testing,attack detection performance is improved.(2)Research under the industrial wireless sensor networks based on WGAN against attack detection model,to overcome the confrontation training can reduce the detection performance of detection model of clean samples challenge,by against the combination of learning and neural network,improve the model structure and the loss function,strengthen the training intensity to strengthen the detection performance of the model.Based on WGAN structure,and the other on the discriminator and the generator add detection classifier with the target classifier respectively,respectively,to fight against the number of training samples,and the counter samples of aggression,to strengthen the training of the detection classifier for counter samples.According to generate against the change of the network structure,adding with the classification of the constraint conditions and against losses to improve the loss function,promote discriminator and generator against the stability of the learning process,to strengthen the robustness of detection classifier.(3)Design a natural gas pipeline IWSN attack monitoring and management system.The model presented in this paper the real-time detection data in a communication network,and will attack detection in real time display on the interface,and to speak against the specific details and recorded data of each sensor,is used to attack was analyzed with the query.To facilitate the update model,simplified model system using a graphical interface against the related operation in the process of training.