Multi-Modal Deception Detection Using Deep Learning

Following decades of study proving that people cannot reliably discern dishonesty, a strong solution combining gestures, auditory characteristics, and micro-expressions is put forth. This multimodal deep learning system records verbal and nonverbal indicators linked to deceitful conduct by examining speech patterns and facial features. Different categorization models combine MLP (Multi-Layer Perceptron) with CNN (Convolutional Neural Network) architectures to process auditory information and LSTM (Long Short-Term Memory) to handle visual features. Accuracy and dependability are increased by combining predictions from many modalities using an ensemble fusion approach. The approach includes ensemble modeling, feature extraction, modality separation, and video preprocessing. With an overall accuracy of $\mathbf{9 4. 2 3 \%}$, evaluations on a carefully selected dataset of trial hearing recordings show excellent performance. The integrated technique provides a non-intrusive and dependable deception detection system by successfully separating honest from misleading remarks. The solution’s potential uses in security, human resources, and law enforcement underscore its significance and promise as a new benchmark for precise lie detection.