Utilizing Artificial Intellegence to Detect Deceit in Videos of Filipinos
Date of Award
Master of Science in Electronics Engineering
Information Systems & Computer Science
Ma. Mercedes T. Rodrigo, PhD
Most of the studies that use artificial intelligence (AI) models to detect lies in videos were done using videos of Caucasians. While previous psychological research has shown that liars generally exhibit distinctive behaviors universal across cultures (universal-cue hypothesis), many actions exhibited by liars are also culture-specific (specific- discrimination hypothesis). This insight implies that the results of the studies that utilize AI models to detect lies in videos may not be might not be generalizable to non- Caucasians. This study builds a model based on machine learning techniques discussed in the literature review and obtains a ROC-AUC of 0.7156 with 6-fold cross-validation. The study finds that audio features are the most effective modality, followed by video features, while micro-expressions on their own are unable to detect lies any better than chance. Since the extraction of audio features is independent of any training data while the extraction of video features and micro-expressions are not, the study concludes that the proposition that the universal-cue and specific-discrimination hypothesis is also applicable to AI holds, but more research is needed to definitively prove this.
(2021). Utilizing Artificial Intellegence to Detect Deceit in Videos of Filipinos. Ateneo de Manila University.