International Journal of Scientific Research and Engineering Development

Automated deception detection remains a challenging problem at the intersection of artificial intelligence, psychology, and cognitive science. Most existing systems only look at one or two types of signals, leaving a lot of behavioural and physiological information on the table. This paper introduces MOSAIC, a multimodal deep learning framework that brings together behavioural, acoustic, linguistic, cardiac, and neural signals in a unified late-fusion architecture for deception analysis. The system processes audio-visual inputs alongside physiological data from ECG and EEG recordings. Each modality produces its own confidence score, and these are combined through a deep neural network to arrive at the final prediction. The EEG sub-model, trained on the DEAP dataset, reaches 76.1% accuracy with an AUC-ROC of 0.750. Across the board, MOSAIC shows that fusing signals from multiple modalities makes the system more robust than any single-channel approach. The framework is also designed with ethical deployment in mind, with interpretability built into the evaluation process.

Sinchana S, Rajashree S Ganganahalli, Muhammed Nidhal,"MOSAIC: Multimodal Orchestration of Signals Across Integrated Channels — A Deep Learning System for Deception Detection" International Journal of Scientific Research and Engineering Development, V9(3): Page(930-935) May-June 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.