A Review on Non-Invasive Multimodal Approaches to Detect Deception Based on Machine Learning Techniques

Abstract

Detecting deception has been investigated by the scientific community for over a century due to its importance in the justice system and homeland security. Attempts to come up with an approach, a system or a framework that serves the purpose of discerning lies from truths has therefore been a major field. This has led researchers to automate the detection process and reduce its invasiveness as much as possible. In addition, machine learning techniques are used with multiple channels of information, known as modals, to increase accuracy in what is known as a multimodal approach. As a result, several research and datasets are currently available, and it could be challenging to identify successful patterns, gaps, and future directions. In this paper, over fifty state-of-the-art publications in the field of deception detection using non-invasive approaches based on machine learning techniques are analyzed after reviewing more than one thousand publications from Scopus, IEEE Xplore, Web of Science, ScienceDirect, and Google Scholar. The work presents the classification techniques and datasets used with their detection performance and finally analyzing the data to draw conclusions. The reported detection accuracy ranges from about 50% to 95% for monomodal approaches based on facial expression, body movement, audio, or thermal imaging. In conclusion, the multimodal approach shows promising results as it reaches a detection accuracy approaching 100%. It outperforms any alternative non-invasive approach, especially when dealing with small datasets, which seems to be the biggest challenge in this field. Future research directions should focus on experimenting with multimodal systems by developing larger datasets as well as implementing classification algorithms that can work with multiple modals effectively.