Micro Motion Aware for Deception Detection

Automated deception detection is important for information integrity and public safety, yet remains challenging because deception-related facial cues are often subtle and difficult to discern. Existing methods remain limited in effectively capturing and modeling these fine-grained facial micro-motion cues. To address this limitation, we propose a Micro-Motion-Aware Automated Deception Detection framework (MMA-ADD), a unified framework that jointly accounts for three key characteristics of deception-related facial micro-motions: localized subtle variations, heterogeneous spatial extents and temporal durations, and structured interactions among facial action units (AUs). First, a Spatio-Temporal Multiscale Embedding (S-TME) employs multiscale 3D convolutions to capture facial micro-motions across heterogeneous spatial extents and temporal durations, establishing a unified representation space for subsequent modeling. Second, a lightweight Micro-Motion Localization Guidance Module (MLGM) accumulates inter-frame differences to guide backbone attention toward facial regions exhibiting subtle motion variations, without introducing an independent motion stream. Third, an Intra-Inter AU Interaction (IIAI) module models both temporal dynamics within individual AU regions and coordinated interactions across different AUs, yielding structured AU-level representations that complement the holistic backbone features. Experimental results on three public datasets, namely DOLOS, Bag-of-Lies, and MU3D, demonstrate that MMA-ADD achieves superior overall performance compared with state-of-the-art methods. The code is available at the MMA-ADD.