INTERNATIONAL JOURNAL FOR INNOVATIVE RESEARCH IN MULTIDISCIPLINARY FIELD

Abstract:    Deception detection has evolved from invasive, mono-modal physiological monitoring (e.g., polygraphy) to non-invasive, automated multimodal frameworks leveraging Artificial Intelligence. Deceptive behavior induces measurable shifts in cognitive load, emotional leakage, and physiological homeostasis. This paper presents a comprehensive, end-to-end framework for Automated Multimodal Deception Detection (AMDD). The system ingests concurrent video, audio, and textual streams, extracting high-fidelity behavioral features: spatial-temporal facial micro-expressions via Video Vision Transformers (ViViT), paralinguistic and acoustic prosody using Hidden-Unit BERT (HuBERT), and semantic/syntactic anomalies via a Large Language Model (LLM). These distinct modalities are combined using a Cross-Attention Fusion network and processed through a generative Chain-of-Thought (CoT) reasoning layer. Evaluated on the benchmark Real-Life Trial and DOLOS datasets, our unified multimodal architecture achieves an F1-score of 0.89, significantly outperforming traditional unimodal baselines. Finally, we address critical ethical, legal, and adversarial constraints inherent to high-stakes algorithmic profiling.