Lies Uncovered: Comparing the Performance of Deep Learning Techniques in Video-Based Deception Detection

The discovery of effective and efficient deception detection techniques in video data has significant implications in various fields such as psychology and law enforcement. With the advancement of computer technology, particularly in the field of deep learning, this research aims to improve the accuracy and efficiency of deception detection in video data. Although previous research has investigated deep learning techniques for deception detection in video data, there is still a need to overcome limitations in the use of real-world datasets, generalization of models, and consideration of their efficiency and effectiveness in practical applications. This research investigated seven deep learning models, including 1D CNN, 2D CNN, 3D CNN, RNN, LSTM, GRU, and Transformer Self Attention, for deception detection in video data. The experimental data comes from real case videos that have been labelled as honest or not based on court decisions. This dataset consisted of 121 videos, of which 61 were labelled as deceptive and 60 as honest. Each video contains 1600-2400 frames, with a total of 241504 data entries processed. It was found that 1-D CNNs were the most promising for detecting deceptive video data, with an accuracy of 0.99 and a testing time of 4.63 seconds, providing a good combination of high accuracy and time effectiveness. The GRU and LSTM also show strong potential, particularly in applications where the trade-off between speed and accuracy is critical. Future research should explore modifying this deep learning method to better suit the context of deception detection by recognizing each frame as a part of a unified video.