Core Concepts
Automated deception detection using multimodal features outperforms traditional methods.
Abstract
The study proposes a voting-based method for automatic deception detection from videos, utilizing audio, visual, and lexical features. Experiments were conducted on two datasets: Real-life Trial Dataset and Miami University Deception Detection Dataset. The proposed solution consists of three models: CNN for images, SVM on Mel spectrograms for audio, and Word2Vec on SVM for manuscripts. Results show significant improvements over state-of-the-art methods in detecting deception across different modalities.
Traditional methods like polygraphs are discussed, highlighting the limitations and psychological impact of such tests. The importance of accurate deception detection in scenarios like job interviews and court trials is emphasized. Previous works in automated deception detection are reviewed, showcasing advancements in using verbal and non-verbal features to enhance accuracy.
The study compares results from different models across image, audio, text, and multimodal components. Best accuracies achieved were 97% for images, 96% for audio, 92% for text on the Real-Life Trial Dataset. Fusion of results from all models yielded an overall accuracy of around 90%. The study concludes that automated deception detection offers a practical solution with improved accuracy compared to traditional methods.
Stats
Best results achieved on images, audio and text were 97%, 96%, 92% respectively on Real-Life Trial Dataset.
Best results achieved on video, audio and text were 97%, 82%, 73% respectively on Miami University Deception Detection Dataset.
Quotes
"By extracting various features from data including visual features such as hand movements and facial features or acoustic features such as tone and pitch or lexical features by analyzing the spoken text... it’s possible to automatically detect deception from videos."
"The proposed solution outperforms previous state-of-the-art models."