An AI-Powered Lie Detection System Using Facial Micro-Expressions

Abstract: Deciding whether a person is telling the truth matters in security screening, forensic questioning, recruitment interviews, and fraud inquiries, yet the available instruments are weak. The polygraph demands attached sensors and a trained operator, and unaided human judges perform only marginally above chance. Micro-expressions, which are involuntary facial movements lasting roughly a twenty-fifth to a fifth of a second, give a contactless cue, but they pass too quickly for people to follow reliably. This paper describes the design and prototype of a video-based deception analysis system built around the Real-life Deception Detection Dataset of 121 courtroom clips. It joins facial-landmark geometry with learned visual features, follows how both change over time with a recurrent network that carries an attention layer, and returns a Truth/Lie probability together with an account of the frames and facial regions that shaped it. The design rests on a short review of six recent studies and on informal consultations with a recruiter and a police officer. The upload, analysis, and explained-output interface of the prototype is working, whereas quantitative validation (accuracy, precision, recall, and F1-score) is still to be carried out, so no performance claim is made. The contributions are a documented architecture, an explicit evaluation protocol, and a discussion of ethical limits that treats the tool as support for a human examiner and not as a verdict.

Keywords: lie detection, deception analysis, facial micro-expressions, MediaPipe, CNN-LSTM, attention mechanism, explainable AI