Research On Contactless Deception Detection Algorithms

With the development of information technology,the ways of communication between people are gradually diversified.It is particularly important to distinguish the authenticity of information in information exchange.Deception detection technology emerges as the times require.Because of the weak operability and low accuracy of contact deception detection technology in the past,automatic non-contact lie detection research is particularly urgent.Therefore,based on the two non-contact deception detection methods of eye movement and speech,machine learning technology is used to explore the feasibility of cheating detection.The specific work is as follows:(1)The background and significance of untouched lie detection are introduced in this paper.The common index parameters in eye movement lie detection and voice lie detection and the role of each feature in lie detection are summarized,and the algorithm models commonly used in voice lie detection by predecessors are expounded.(2)Some common theoretical foundations of untouched lie detection are discussed.Two lie detection paradigms,GKT and CQT,are briefly introduced.Cognitive meanings of gaze,eye jump,blink and pupil size indicators in lie detection are introduced.The computational processes of acoustic features in MFCC,F0 and PLP3 speech lie detection fields are analyzed.Finally,feature selection algorithms: PCA,SFFS and variance analysis are introduced.(3)The experimental process and configuration of eye movement experiment are introduced.Machine learning is applied to eye movement lie detection.Five models,GMM,SVM,ANN,decision tree and Fisher classifier,are used respectively.Eye movement indicators combined 32 eye movement indicators to analyze the difference between deceivers and nondeceivers.And the feasibility of eye movement lie detection is verified by the comparative experiments of different feature sets on different machine learning models.The results show that the highest recognition rate of SVM model is 78.13% on Ren5 dataset,which shows the feasibility of eye movement lie detection.(4)Deep learning method is applied to speech lie detection,and a convolutional cyclic neural network structure integrating attention mechanism is proposed.To verify the validity of the model for speech lie detection,comparative experiments were conducted on CSC corpus using CNN + spectral features,BiLstm + frame-level features and ConvBilstm + Attention + frame-level features.The results show that the recognition results of the model are improved from 62.15% to 68.99% compared with single convolution or cyclic neural network,and 64.00% compared with traditional machine learning algorithm,which shows the effectiveness of the algorithm.