Facialcuenet: An Interpretable Deception Detection Model for Criminal Interrogation Using Facial Expression

Abstract

We propose a deception detection method using non-contact biometric information based on deep learning technology. We aimed to develop an algorithm applicable to actual criminal interrogation. For this purpose, we developed a data acquisition protocol to satisfy the conditions similar to the actual investigation environment. Public and collected datasets, according to the ‘DDCIT dataset’ protocol, were used to evaluate the algorithm. The videos in the dataset were analyzed based on a developed deep neural network called ‘FacialCueNet,’ with an attention module applied to the spatial-temporal domain, and action unit frequency, symmetry, gaze pattern, and micro expression were extracted as facial cue features and inserted into the network. As a result, the mean deception detection F1-score using the DDCIT dataset was 81.22%, and evaluation accuracy against the public database was 88.45%. We also presented interpretive results of deception detection by analyzing the influence of spatial and temporal factors.

Keywords: Deep Learning, Attention network, Investigation, Deception detection