[Paper Review] GenLie: A Global-Enhanced Lie Detection Network under Sparsity and Semantic Interference

Z. H. Zhang, Yao Liu+6 more|arXiv (Cornell University)|Mar 14, 2026

Deception detection and forensic psychology0 citations

TL;DR

GenLie introduces a local–global framework for video-based lie detection that learns sparse, discriminative cues while suppressing identity-related noise via adversarial decorrelation and a triplet loss, achieving state-of-the-art results across three benchmarks.

ABSTRACT

Video-based lie detection aims to identify deceptive behaviors from visual cues. Despite recent progress, its core challenge lies in learning sparse yet discriminative representations. Deceptive signals are typically subtle and short-lived, easily overwhelmed by redundant information, while individual and contextual variations introduce strong identity-related noise. To address this issue, we propose GenLie, a Global-Enhanced Lie Detection Network that performs local feature modeling under global supervision. Specifically, sparse and subtle deceptive cues are captured at the local level, while global supervision and optimization ensure robust and discriminative representations by suppressing identity-related noise. Experiments on three public datasets, covering both high- and low-stakes scenarios, show that GenLie consistently outperforms state-of-the-art methods. Source code is available at https://github.com/AliasDictusZ1/GenLie.

Motivation & Objective

• Motivate the need for robust lie detection from subtle, sparse visual cues in videos.
• Propose a local–global modeling strategy to extract sparse cues while reducing identity and contextual noise.
• Develop an end-to-end framework with redundancy-aware sampling, task-driven re-embedding, adversarial decorrelation, and a discriminative loss.
• Demonstrate state-of-the-art performance across both low-stakes and high-stakes deception datasets.

Proposed method

• Segment videos into equal-length clips and select top-K informative frames per segment to suppress redundancy.
• Use a frozen VideoMAEv2 encoder to extract segment-level features and obtain a global video embedding via a lightweight MLP.
• Apply adversarial speaker-decorrelation to produce identity-invariant representations through a GRL-based auxiliary classifier.
• Incorporate a video-level triplet loss to enforce intra-class compactness and inter-class separation.
• Train with a joint objective L_total = L_cls + alpha L_id + beta L_tri to balance discrimination and debiasing.

Fig. 1 : Core challenge in video-based lie detection: subtle and transient cues are often overwhelmed by noise and contextual bias.

Experimental results

Research questions

• RQ1 Can a local–global modeling approach effectively capture sparse deceptive cues while mitigating identity-related noise in video deception detection?
• RQ2 Does adversarial speaker-decorrelation improve cross-speaker generalization without sacrificing deception discriminability?
• RQ3 Is a video-level triplet loss beneficial for shaping a more discriminative embedding space across deception labels and speakers?
• RQ4 Which frame selection strategy yields the most robust performance across datasets with varying cue salience?

Key findings

• GenLie achieves state-of-the-art performance across three deception benchmarks (MDPE, Real-Life Trial, SEUMLD).
• Ablation shows semantic re-embedding is crucial for performance, with speaker decorrelation contributing notably on Real-Life Trial.
• Uniform frame sampling provides robust performance compared to heuristic sampling strategies.
• Temporal segmentation helps preserve sparse cues, and triplet loss strengthens discriminability.
• GenLie excels in both low-stakes and high-stakes settings, demonstrating strong generalization across identities.

Fig. 2 : Architecture of the proposed GenLie framework.