In recent years, the occurrence of numerous crimes and issuance of incorrect judicial verdicts have pro‎mp‎ted researchers to look for a scientific method to distinguish between guilty and innocent people. Therefore, different deception detection methods were presented in the form of concealed information test, in which the level of deception was detected by examining the body's physiological responses such as temperature, heart rate, breathing rate, blood pressure, etc. Examining the level of deception based on physiological responses had a low level of confidence due to the fact that it was easy for the subject to control their responses. This lead to popularization of deception detection methods based on brain signals, in which by applying stimuli both related and unrelated to the crime, the changes in the person's brain signals are examined. Based on that, a decision is made on whether the subject is guilty or innocent. In this thesis, the database that was collected by Gao in 2014 was used, which contains the information of 15 guilty and 15 innocent subjects. In the pre-processing stage, the EEG signal was epoched, baseline was corrected and EOG artifacts were removed. Then, by using the appropriate embedding dimension and time lag, the signals were transferred from the time domain to the phase space. 5 different methods including recurrence plot, cross recurrence plot, order pattern recurrence plot, order pattern cross recurrence plot and modified multiscale entropy were used for feature extraction. Classification was investigated using 2 different approaches. In the first approach, information retrieved from different channels was considered as different features and the appropriate features were selected using the statistical dependency function and separated by 4 classifiers: support vector machine, k-nearest neighbor, random forest and linear classifier. In the second approach, the information of different channels was considered as different observations and appropriate features were selected by the statistical dependency function. Then, by applying a spatial filter and providing an optimization model, the features were transferred to a new space and classified by the same classifiers as the first approach. The results of this research indicate improvement in classification in the second approach compared to the first. It was observed that the features of order pattern recurrence plot and modified multiscale entropy have a higher performance in differentiating guilty and innocent subjects and cause more classification accuracy than other features.

امير اخوان

فرزانه شايق، احسان روحاني