Abstract: PURPOSE: Facial expression, as a potential non-verbal cue, is of great significance for lie detection. This study aims to compare the performance of machine learning models using happy and fearful expressions as indicators for distinguishing lies from truth.
METHODS: By designing a deception game experiment, we recorded videos of subjects lying and telling the truth. Using the facial analysis software OpenFace, we extracted two features of happy expressions (AU6 and AU12) and six features of fear expressions (AU01, AU02, AU04, AU05, AU20, AU26). These features are fed into the machine learning software WEKA for classification.
RESULTS: When using happy expression features for classification, our machine learning model has an accuracy of 77.80% (Random Forest), 77.99% (IBK), and 75.14% (Bagging) in distinguishing lies from truth. When using fear expression features for classification, the accuracy rate was significantly increased to 94.03% (Random Forest), 92.78% (IBK), and 89.36% (Bagging).
CONCLUSIONS: Fearful expressions are more effective than happy expressions as indicators for distinguishing lies from truths.

Key words: fear, happy, machine vision, deception detection, facial expression