Department of Computer Engineering
CS 590 SEMINAR
Deceit Detection From Behavioral Cues
Computer Engineering Department
Lies and deception are social phenomena both in our private and professional lives. On average, human accuracy of distinguishing between lies and truthful statements is 54% that is just above the random guess. Therefore, high performance deceit detection methods are necessary in cases such as interrogation and evaluating the testimonies of witnesses in courtrooms. In this thesis, the aim is to develop an easy-to-use deception detection model based on visual analysis from the videos of people recorded during the interrogation for both high-stake and low-stake scenarios. In the analysis, spatio-temporal convolutional neural networks that take into account temporality along with spatial features such as facial expressions, action units and micro-expressions are utilized to catch the subtle facial cues formed by the subconscious. Ultimately, given a query video without a specific length limitation, the model determines whether the person in the video is telling the truth or not. The widespread effects of the model may include assisting decision makers in interrogations, where the truthfulness of the discourse is of great importance, and in job interviews evaluating how honestly candidates behave.
DATE: 16 December 2019, Monday @ 15:40