Reproduce “DeceptionX: From Multimodal Evidence to Explainable Deception Detection” on OpenResearch

About this paper

Jiayu Zhang, Shuo Ye, Jiajian Huang, Yawen Cui, Taorui Wang, Wei Xia, Zeheng Wang, Haowen Tang and 3 more

DeceptionX, a multimodal framework for explainable deception detection, was developed to mimic human reasoning by creating the DeceptChain dataset and utilizing a three-stage progressive training with a Discrepancy-Aware Redundancy Elimination (DARE) strategy. The framework achieved superior accuracy and F1 scores (e.g., 70.12% Acc, 73.26% F1 on DOLOs) compared to existing black-box and MLLM models, while also providing high Semantic Evidence Alignment for its generated reasoning paths.

Key insights

• High-quality, structured chain-of-thought reasoning data is essential for enabling MLLMs to perform complex, interpretable behavioral analysis tasks such as deception detection.
• A progressive, multi-stage training paradigm is effective in equipping MLLMs with the ability to first align multimodal cues, then reason causally, and finally optimize for both accuracy and logical plausibility.
• Dynamically focusing model attention on diagnostically critical segments and cross-modal incongruences via redundancy elimination significantly improves both detection performance and the quality of generated explanations.

Summary by alphaXiv.