Noncooperative Signal Analysis Via Deep Learning

Non-cooperative signal analysis technology has been widely used in electronic in-formation warfare and other fields.Based on this technology,to decipher enemies in-telligence or interfere with their communications,non-cooperative receiver utilizes inter-cepted signal to obtain information of the transmitter.In the modern digital communi-cation systems,data is usually transmitted in units of frames.The cooperative receiver will first adopt an appropriate algorithm and combine the information of the frame struc-ture to obtain frame synchronization.However,these are unknown to the non-cooperative receiver,so it needs to utilize intercepted signals to identify the frame structure.After ob-taining the frame synchronization,if the non-cooperative receiver wants to further get the information of the data part in the frames,it needs to identify the channel codes adopted by the intercepted signal.This thesis proposes blind recognition algorithms based on deep learning for the two problems.This thesis first considers the blind recognition of the frame structure under the Rayleigh fading channel.For the identification of frame length and sync word length,this paper combines recurrent neural network(RNN)with the window-based correlation function method and the periodic sampling-based mean function method,and proposes a frame length recognizer based on window-based recurrent neural network(WRNN)and sync word length recognizer based on sample-based recurrent neural network(SRNN).For the identification of delay,we propose a correlation function method based on window peri-odic sampling from the perspective of asymptotic analysis,and then combine it with RNN to propose the delay recognizer based on sample-window-based recurrent neural network(SWRNN).The three recognizers need near zero prior information and only require sev-eral frames.The simulation results show that the proposed recognizers have strong gen-eralization to the testing samples that are not in the training set,and compared with the RNN recognizer without utilizing conventional algorithms,the number of training sam-ples required is greatly reduced.Then,this thesis considers the blind recognition of the type and the encoding parame-ters of channel codes from the noisy signals.Specifically,based on the RNN,the attention mechanism,and the residual network(Res Net),three universal recognizers are proposed to identify the type,rate,and length of the target channel codes,with a training set gen-erated by a small portion of all the possible code parameters.The proposed architectures need near zero prior knowledge about the target channel code,and only require the length of the received signal to be dozen times of the codeword length.Numerical experiments show that the proposed deep learning methods own strong generalization to identify chan-nel codes from the testing samples not generated by the encoding parameters utilized for the training set.