FaceVeritas: Real-Time AI-Based Lie Detection System

Abstract

Automated deception detection remains an open challenge in behavioral AI, with traditional polygraph systems constrained by invasiveness, inconsistent reliability, and limited scalability. This paper presents FaceVeritas, a real-time, non-invasive lie detection framework that identifies deceptive behavior exclusively through facial micro-expression analysis using computer vision, without reliance on speech, physiological sensors, or text input. The system captures live video at 30FPS, extracts 468 three-dimensional facial landmarks via MediaPipeFaceMesh, and computes seven behavioral features per frame — Eye Aspect Ratio (EAR), blink rate, Mouth Openness Ratio (MOR), eyebrow lift, head yaw (?yaw), head pitch (?pitch), and normalized face distance (rface). Raw features are temporally stabilized using Exponential Moving Average (EMA, ? = 0.2) to suppress landmark jitter while preserving genuine micro-expression transients. A supervised Random Forest classifier (100 trees) trained on the Bag-of-Lies and Real-Life Trial datasets generates a continuous deception probability score per frame. Experimental evaluation on a held-out test set of 200 samples achieves 78.5% accuracy, 77.1% precision, 81.0% recall, 79.0% F1-score, and 92ms per-frame inference on a standard Intel i5 CPU without GPU acceleration — sufficient for real-time deployment. Blink rate (28.3%) and head yaw (24.1%) emerge as the strongest feature discriminators. FaceVeritas outperforms voice stress analysis by 13.7 percentage points and exceeds manual FACS annotation by 7.3 percentage points, while requiring only a standard consumer camera. The interpretable Random Forest architecture [11] addresses a critical gap in forensic applicability compared to black-box deep-learning alternatives.

Introduction

The text presents FaceVeritas, a real-time, camera-based deception detection system designed to overcome the limitations of traditional methods like polygraphs, which rely on physiological signals that are not specific to lying and are often unreliable. Instead, FaceVeritas focuses on facial micro-expressions—brief, involuntary facial movements that reveal concealed emotions.

Using computer vision and machine learning, the system extracts seven key behavioral features from facial landmarks, including eye aspect ratio (EAR), blink rate, mouth openness, eyebrow movement, head pose, and face distance. These features are smoothed using an Exponential Moving Average (EMA) and analyzed using an interpretable Random Forest classifier, allowing transparency in decision-making—important for forensic use.

The system operates through a six-step pipeline: video capture, face detection, feature extraction, smoothing, classification, and result visualization. It runs in real time on standard consumer hardware without requiring additional sensors or internet connectivity.

Experimental results show that FaceVeritas achieves around 78.5% accuracy, with strong recall and an AUC of 0.843, indicating good performance in detecting deception. Feature analysis reveals that blink rate and head movement are the most significant indicators of deceptive behavior.

Compared to other methods, FaceVeritas balances accuracy, real-time performance, low cost, and interpretability, outperforming simpler camera-based approaches while avoiding the complexity of multimodal or sensor-based systems.

Conclusion

This paper presented FaceVeritas, a real-time AI-based deception detection system built on facial micro-expression analysis using MediaPipeFaceMesh, a seven-dimensional behavioral feature set (EAR [4], blink rate, MOR, eyebrow lift, head yaw/pitch, normalized face size), EMA temporal smoothing (? = 0.2), and Random Forest [11] classification. The framework operates on a single standard camera without physiological sensors, achieves 78.5% accuracy at 92ms classification latency on a consumergrade CPU, and provides per-feature MDI importance rankings that satisfy forensic interpretability requirements. Comparative evaluation confirms superiority over voice stress analysis and manual FACS [3] coding, with competitive accuracy against GPU-accelerated deep learning at a fraction of the hardware cost. Planned future directions include: 1) Deep feature learning: CNN-LSTM encoder replacing hand-crafted features for automatic temporal pattern discovery, with knowledge distillation to maintain interpretability. 2) Subject calibration: Two-minute neutral baseline calibration to personalize EAR and blink thresholds per subject, reducing individual variability false positives. 3) Dataset expansion: Collection of demographically diverse, cross-cultural deception corpora under varied naturalistic lighting conditions. 4) Edge deployment: Model quantization (INT8) and TensorFlow Lite export for mobile forensic deployment on resourceconstrained hardware. 5) Multimodal extension: Optional voice prosody channel integration when audio is available, with modality-dropout training to preserve camera-only fallback. 6) Ethical audit framework: Automated fairness evaluation across demographic subgroups and certified deployment protocols for law enforcement and corporate use.

References

[1] A. Vrij, Detecting Lies and Deceit: Pitfalls and Opportunities, 2nd ed. Chichester, UK: Wiley, 2015. [2] P. Ekman, “Lie catching and microexpressions,” in The Philosophy of Deception, C. Martin, Ed. Oxford, UK: Oxford University Press, 2009, pp. 118–133. [3] P. Ekman and W. V. Friesen, Facial Action Coding System (FACS): A Technique for the Measurement of Facial Action. Palo Alto, CA: Consulting Psychologists Press, 1978. [4] T. Soukupova and J.´ Cech, “Real-time eye blink detection using facial landmarks,” in? Proc. 21st Comput. Vision Winter Workshop (CVWW), RimskeToplice, Slovenia, Feb. 2016. [5] M. Rahmani, M. H. Ghazvini, H. Ahmadi Zanjani, and A. Mousavi, “Automated systems for detection of deception in written texts and facial behavior: A review,” Sensors, vol. 24, no. 8, Art. no. 2630, Apr. 2024. [6] H. Guo, S. Chen, and H. Zhang, “DOLOS: A large-scale deception detection dataset with diverse modalities,” in Proc. IEEE/CVF CVPR, Vancouver, BC, Canada, Jun. 2023, pp. 1630–1639. [7] O. Shalash, D. Mohamed, and A. Elbarkouky, “Lie detection based on facial micro-expressions, heart rate and blood volume pulse,” Inventions, vol. 10, no. 1, Art. no. 9, Jan. 2025. [8] S. Tsiamyrtzis, J. Dowdall, D. Shastri, I. T. Pavlidis, M. G. Frank, and P. Ekman, “Imaging facial physiology for the detection of deceit,” Int. J. Computer Vision, vol. 71, no. 2, pp. 197–214, Feb. 2024. [9] C. Lugaresiet al., “MediaPipe: A framework for building perception pipelines,” arXiv preprint arXiv:1906.08172, Jun. 2020. [10] V. Perez-Rosas, M. Abouelenien, R. Mihalcea, and M. Burzo, “Deception detection using real-life trial data,” in´ Proc. 2015 ACM Int. Conf. Multimodal Interaction (ICMI), Seattle, WA, USA, Nov. 2015, pp. 59–66. [11] L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. [12] D. Avola, M. Bernardi, L. Cinque, F. D’Amico, and P. Napoletano, “Lie detection through automated facial expression analysis,” in Proc. Int. Conf. Image Analysis and Processing (ICIAP), Trento, Italy, Sep. 2019, pp. 143–154.

Copyright

Copyright © 2026 Sahil Mali, Rahul Chakane, Suresh Pathare, Supriya Baviskar, Prof. Vijayalaxmi Tadkal. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.