GenLie: A Global-Enhanced Lie Detection Network under Sparsity and Semantic Interference
New computer vision network identifies subtle deceptive cues by separating identity noise from genuine signals.
A research team led by Zongshun Zhang and seven other authors has introduced GenLie, a novel AI system for video-based lie detection that tackles the fundamental problem of sparse and subtle deceptive signals. The network employs a dual approach: capturing fleeting, discriminative cues at the local level while using global supervision to filter out identity-related noise and contextual variations that typically overwhelm genuine deception indicators. This global-enhanced architecture allows the model to focus on the brief, often imperceptible visual tells that characterize lying behavior.
GenLie was rigorously tested across three public datasets representing diverse scenarios, from high-stakes interrogations to casual conversations, and consistently outperformed existing state-of-the-art methods. The research paper has been accepted for presentation at the prestigious IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 2026, indicating its technical significance. The team has made the source code publicly available, encouraging further development and application in fields ranging from security screening to psychological research.
- Uses local feature modeling with global supervision to capture sparse deceptive cues
- Successfully suppresses identity-related noise that typically interferes with detection
- Outperforms existing methods on three public datasets covering varied scenarios
Why It Matters
Advances automated deception detection for security, hiring, and research while addressing fundamental technical challenges.