High stakes can be stressful whether one is telling the truth or lying. However, liars can feel extra fear from worrying to be discovered than truth-tellers, and according to the “leakage theory,” the fear is almost impossible to be repressed. Therefore, we assumed that analyzing the facial expression of fear could reveal deceits. Detecting and analyzing the subtle leaked fear facial expressions is a challenging task for laypeople. It is, however, a relatively easy job for computer vision and machine learning. To test the hypothesis, we analyzed video clips from a game show “The moment of truth” by using OpenFace (for outputting the Action Units (AUs) of fear and face landmarks) and WEKA (for classifying the video clips in which the players were lying or telling the truth). The results showed that some algorithms achieved an accuracy of >80% merely using AUs of fear. Besides, the total duration of AU20 of fear was found to be shorter under the lying condition than that from the truth-telling condition. Further analysis found that the reason for a shorter duration in the lying condition was that the time window from peak to offset of AU20 under the lying condition was less than that under the truth-telling condition. The results also showed that facial movements around the eyes were more asymmetrical when people are telling lies. All the results suggested that facial clues can be used to detect deception, and fear could be a cue for distinguishing liars from truth-tellers.

No abstract

No abstract

No abstract

The leakage theory in the field of deception detection predicted that liars could not repress the leaked felt emotions (e.g., the fear or delight); and people who were lying would feel fear (to be discovered), especially under the high-stake situations. Therefore, we assumed that the aim of revealing deceits could be reached via analyzing the facial expression of fear. Detecting and analyzing the subtle leaked fear facial expressions is a challenging task for laypeople. It is, however, a relatively easy job for computer vision and machine learning. To test the hypothesis, we analyzed video clips from a game show 'The moment of truth' by using OpenFace (for outputting the Action Units of fear and face landmarks) and WEKA (for classifying the video clips in which the players was lying or telling the truth). The results showed that some algorithms could achieve an accuracy of greater than 80% merely using AUs of fear. Besides, the total durations of AU 20 of fear were found to be shorter under the lying condition than under the truth-telling condition. Further analysis found the cause why durations of fear were shorter was that the duration from peak to offset of AU20 under the lying condition was less than that under the truth-telling condition. The results also showed that the facial movements around the eyes were more asymmetrical while people telling lies. All the results suggested that there do exist facial clues to deception, and fear could be a cue for distinguishing liars from truth-tellers.