Construction And Analysis Of A Database Of "Judged" Expressions In Gaming

Deception and deception detection have been extensively studied in the field of psychology.Researchers have designed various experimental tasks to induce deceptive behavior,such as disguise tasks,lying tasks,and stealing tasks.They have also used neuroimaging techniques,physiological indicators,and facial expression analysis to explore the physiological and neural mechanisms of deception.However,there are relatively few studies on deception and deception detection in a gaming context,where both deceptive and honest behavior are scrutinized by the opponent,and individuals experience significant pressure due to the competitive nature of the game,leading them to conceal their true thoughts and intentions.Currently,studies on the deception expression database mainly use a single-person paradigm to collect video data of subjects engaging in deceptive behavior.These data typically cover facial expression information after the subject engages in deceptive behavior,but do not include video data from the "trial" where the subject is judged.Therefore,in this study,we used a deception game paradigm to design a two-person game in which the subject needs to win by deception.We collected video data of subjects being "tried" after engaging in honest or deceptive behavior during the game.In this process,due to the double pressure brought by the anxiety of deception and fear of being scrutinized by the opponent,the subject’s true thoughts may be revealed through facial expressions.We named the video clips collected during this stage the Gaming is’ tried’ database(GTD).After establishing the the Gaming is ’tried’ database database,the video materials in the database were analyzed using Openface software to extract facial coordinate data,and Python software was used to integrate and calculate the data to obtain 20 distance indicators.Based on this,three analysis methods were employed,including distance indicator analysis,facial symmetry analysis,and facial action unit feature analysis,to explore the differences in facial expression features between deceptive and honest behaviors.The first analysis involved using Openface software to extract facial coordinate data from the video materials and then using Python software to integrate and calculate the data to obtain 20 distance indicators,including Ld1,Ld2,Ld3,Md1,Md2,Md3,Md4,Lfd,Rd1,Rd2,Rd3,Mfd,Lpupil＿d,Rpupil＿d,ratio＿r,ratio＿l,ratio＿m,Lpupil＿d,Rpupil＿d,and blink＿mean.Suitable distance indicators were selected,and paired sample t-tests were performed on the distance indicators under honest and deceptive contexts using SPSS software to investigate whether significant differences existed between these distance indicators in the two contexts.The second analysis involved analyzing facial symmetry by selecting the distance indicators ld1 and rd1 for each participant under both honest and deceptive contexts.The wavelet coherence function in Matlab software was used to calculate the coherence of facial movements,which indicates higher symmetry.Paired sample ttests were also conducted using SPSS software to investigate whether facial symmetry differed between honest and deceptive contexts.The third analysis involved analyzing facial action units(AU),specifically,selecting 17 AU features including AU01,AU02,AU04,AU05,AU06,AU07,AU09,AU10,AU12,AU14,AU15,AU17,AU20,AU23,AU25,AU26,and AU45.Paired sample t-tests were conducted using SPSS software to investigate which features showed significant differences between honest and deceptive contexts.The results indicated that four features,namely ratio＿m,blink＿mean,AU07,and AU14,showed significant differences.Using these features,machine learning was performed in Matlab software using four algorithms,including decision trees,support vector machines,K-nearest neighbors classification,and bagging trees.The bagging tree algorithm had the highest accuracy,reaching 72%,which was significantly higher than random guessing.The model performance was evaluated based on AUC,precision,recall,F1 score,and specificity indicators,and the bagging tree model had the best performance.This study used the deception game paradigm to study the facial expression features of participants during the "judgment" stage under honest and deceptive contexts.The obtained data was used to construct a dataset and a machine learning model,which improves the accuracy of deception detection using artificial intelligence.This research approach helps to comprehensively understand the relationship between deceptive behavior and facial expressions and provides valuable data resources for related fields of research.