Lie detection - The Technical Perspective

Dinges, L., Fiedler, M. A., Al-Hamadi, A., Hempel, T., Abdelrahman, A., Weimann, J., Bershadskyy, D., & Steiner, J. (2024). Exploring facial cues: automated deception detection using artificial intelligence. Neural Computing and Applications, 1–27. https://doi.org/10.1007/S00521-024-09811-X/FIGURES/17

Dinges, L., Fiedler, M.-A., Al-Hamadi, A., Abdelrahman, A., Weimann, J., & Bershadskyy, D. (2023). Uncovering Lies: Deception Detection in a Rolling-Dice Experiment. In G. L. Foresti, A. and Fusiello, & E. Hancock (Eds.), Image Analysis and Processing -- ICIAP 2023 (pp. 293–303). Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-43148-7_25

In this joint work between economists and neuro-information scientists, we explore the potential of automated deception detection using artificial intelligence (AI), focusing on facial cues such as gaze, head pose, and facial expressions. By employing convolutional neural networks (CNNs) we evaluate deception detection across various datasets, including our own data set building on the seminal work of Fischbacher & Föllmi-Heusi (2013) involving dice rolls.

Performance Variability: Facial expressions consistently outperformed gaze and head pose in detecting deception. Combining multiple modalities enhanced accuracy, though results varied depending on dataset contexts.
Challenges of Low-Stakes Contexts: Low-stakes scenarios like the dice-rolling experiment presented greater difficulties for detection compared to high-stakes datasets such as courtroom trials.
Dataset-Specific Models: Results highlighted the importance of context-specific training data. Models trained on one dataset often underperformed when applied to another, emphasizing the contextual nature of deceptive behavior.
Potential Applications: Automated systems show promise for diverse fields like virtual sales, healthcare, and security, offering a non-invasive alternative to polygraph tests.

In total, we show AI’s potential to improve deception detection but stress the need for more extensive, diverse datasets and scenario-specific approaches to optimize accuracy and applicability.

Page updated

Google Sites

Report abuse