An AI-based lie detector

• Main Author: Wang, Yuyao
• Other Authors: Alex Chichung Kot
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Computer and Information Science; Engineering
• Online Access: https://hdl.handle.net/10356/182190
• Institution: Nanyang Technological University

Deception detection is a very meaningful research direction. Traditional deception detection methods, such as those that rely on biosignals, are often problematic due to their limited accuracy, invasiveness, and susceptibility to manipulation. In contrast, deep learning-based deception detection offers a more sophisticated and flexible al-ternative. These models can analyze language patterns, voice modulations, and mul-timodal behavioral data to capture subtle clues that traditional methods may miss. In addition, deep learning models can be effectively scaled to real-world applications and better meet the requirements of modern ethical and practical standards. This makes them the best choice for reliable and robust deception detection. Data in the real world typically includes multiple sources and domains. Therefore, de-signing a fair and unbiased AI deception detection model, that is, training a model with good generalization ability that can learn universal features between different races and cultural backgrounds for deception detection, has become crucial. This project collects a deception detection dataset containing three Asian ethnic groups, including multiple modalities (audio and video) and multiple languages (English, Chinese, Malay, Hindi). Compared with previous public deception detec-tion datasets, this dataset performs better in terms of sample diversity, gender bal-ance, and modality richness. This project focuses on the domain generalization (DG) of cross-racial deception detection in audio modality. Domain generalization (DG) refers to allowing deep learning models to acquire knowledge from multiple related source domains and apply it to unknown domains, thereby improving the model's capacity for generalization when facing out-of-distribution data (OOD). Experi-ments are deployed in English (as a common language) and respective Native Lan-guages. Many domain generalization methods such as gradient reversal, contrastive learning, temperature scaling, etc. are implemented. Finally, a combined loss func-tion combining contrastive learning with focal loss is proposed. This method im-proves the average domain generalization performance by 4.28% compared with the experimental results of the baseline.