Application of Artificial Neural Network in Deception Detection Based on Visual Cues

As deception is a component of our daily social interactions, how to automatically determine a lie has been a topic of great interest to the research community. Faced with the explosive growth of social networks, it is crucial to assess whether the information obtained is accurate, especially in terms of security and social order. According to studies, no clue or clue pattern is specific to deception, although there are clues specific to emotion or cognition that refer to falsification. In this paper, we address the identification of deceit by using single-modal approaches and present a real-time detection system based on Deep Learning. We focus primarily on non-verbal modalities, particularly visual appearance, gaze, and action unit features. We employ LSTM-based networks to address the issue, particularly in two ways: an end-to-end CNN combined LSTM model, called LRCN, and a method that extracts gaze and action unit features separately, then brings them to the LSTM model for learning and classification. The Bag-of-Lies dataset is utilized to evaluate the proposed methods. We achieved classification accuracies between 61% and just over 70%, with the LRCN model ranking highest and contributing to our detection program.