Research On Indoor Device-free Intrusion Detection And Target Classification Method Based On CSI

In recent years,with the rapid development of WLAN technology and the widespread deployment of WiFi worldwide,it has been found that WiFi can be used not only for communication but also for environment sensing.The device-free sensing technology based on WiFi is developed on the basis of the existing WLAN facilities.Compared with the active sensing technology,The device-free sensing technology can use the change of wireless signal in the wireless deployment environment to determine whether an intrusion has occurred and identify the intruder's characteristic information,which has high research significance and commercial value.However,the accuracy of existing device-free intrusion detection and target classification methods has not yet reached commercial requirements,and is still in the research stage.The thesis makes an in-depth study on how to judge the invasion more accurately and classify the characteristics of the personnel after the invasion.main tasks as follows:For device-free intrusion detection systems,most of the existing solutions use a threshold-based method in the judgment stage.The threshold of this method is fixed and cannot adapt to the environmental changes in daily use scenarios,which will bring a high false alarm rate and the problem of poor robustness.In order to solve the above problems,this thesis proposes a KNN intrusion detection method based on CSI dynamic variance.This method uses the KNN algorithm in the decision stage,and uses the similarity metric value as an adaptive intrusion detection discrimination condition to make the pending data The discriminant result depends on how similar it is to the data in the existing fingerprint database,and the data in the fingerprint database is updated as the environment changes,so using the KNN to make decisions can adapt to changes in the environment.This significantly improves the robustness of the system and reduces the false alarm rate.The thesis validates the effectiveness of the proposed method by conducting experiments in indoor environments with rich multipaths and corridors with few multipaths.Experimental results show that the intrusion detection rate of this method reaches 99%,which is 8% higher than the threshold-based method,and the false alarm rate is reduced by 24%.In the course of the experiment,in order to be closer to the actual application,we need to transmit the data of the experimental equipment between different private subnets point-to-point,but the existing scheme is difficult to implement due to the limitations of NAT and firewall devices in the network architecture in reality.In order to solve this problem,an Internet-wide serial port transparent transmission and program update method is proposed.The data is segmented and packaged through the MQTT protocol,bypassing the characteristics of NAT and firewall devices rejecting incoming connections,and realizing different private Point-to-point transmission of data between subnets.For device-free target classification,the existing scheme has two problems.One is to use the RSSI as the base signal.RSSI is just the superposition of multipath signals.To improve the classification accuracy,the number of wireless transceivers must be increased.This will lead to an imbalance between accuracy and cost;the second is manual feature extraction,which will inevitably bring greater work overhead,and it is difficult to promote the use in real scenarios.In view of the above problems,this thesis proposes a device-free target classification method based on BP neural network.Improvements have been made in two aspects: one is to use CSI instead of RSSI as the base signal,CSI is the physical layer information,a pair of wireless signal transceiver devices can obtain enough information,so that more fine-grained sensing can be achieved at low cost;The second is to use the BP neural network instead of manually extracting features.The BP neural network has the ability to learn signal features autonomously through network training,avoiding manual intervention in the system training process and reducing work overhead.The thesis carried out experimental verification of height classification in an office environment.The experimental results show that the accuracy of height classification can reach more than 92%.