A Study Of Gesture Recognition Technique Based On Deep Learning In Wi-Fi Environment

Recent advances in wireless technology have greatly expanded the use of Wi-Fi from providing laptop connectivity to connecting mobile and smart devices to the Internet and home networks.Different gesture changes within the Wi-Fi signal coverage can cause significant wireless channel interference.Therefore,different gestures can be identified based on the correlation between the dynamic changes of Channel State Information(CSI)and gesture movements obtained in commercial Wi-Fi devices.Users can authenticate through gestures to customize the personalized services of smart devices.At the same time,intruders can also mimic gestures that characterize user behavior to control Internet of things devices or access private data.Therefore,privacy protection and privacy attacks in gesture awareness constitute two important components of cyberspace security.Attackers can identify user gestures through changes in wireless signals and steal private data.However,Wi-Fi signals arriving at the receiving device often carry information about specific environments and users,and attackers often need to retrain the model in the new environment.At the same time,in scenarios such as password input,it is difficult to steal a large number of samples,which may lead to the problem of small samples with a small number of samples.Therefore,this dissertation addresses the problems of limited samples and the difficulties of gesture recognition during privacy attacks caused by changes in the environment and user,with the goal of constructing gesture recognition techniques that do not require target domain data,and providing new attack samples for the study of cyberspace security.The main research work and contributions of this dissertation are as follows:(1)A framework for expanding a small-sample Wi-Fi gesture recognition network is proposed.After a series of pre-processing of CSI data with interpolation,outlier removal,phase correction,wavelet denoising,band-pass filtering and static component removal,the amplitude and phase,which are more sensitive to fine-grained gestures,are extracted as hybrid features.In this dissertation,we propose a novel adjunct generative adversarial network to perform sample augmentation,using Wasserstein distance instead of Jensen-Shannon scatter to solve the classification accuracy problem of generated data and improve the stability and speed of training.Fifty American Sign Language actions were collected as a dataset in five different environments and nine users of different heights and weights.The sample augmentation network can effectively improve the gesture recognition accuracy in the case of insufficient sample size.(2)A domain-adaptive Wi-Fi gesture recognition network framework across environments and users is proposed.In this dissertation,we propose a novel recurrent generative adversarial network to realize the style transformation from source to target domains for cross-environment and cross-user gesture recognition.Specifically,DenseNet is utilized in the generator network structure to speed up the network convergence and enhance the feature transfer capability.L1 loss and feature-level loss are added to the loss function to penalize the difference between synthetic CSI feature data and real data,constrain the feature-level style similarity,and improve the network domain adaptive capability.The experimental results show that the domain adaptive network outperforms other state-of-the-art methods with the highest accuracy of 91.67%and 91.75%for cross-environment and cross-user gesture recognition.