Department of Computer Engineering
DLC: Deceit Localization and Comparative analysis model based on novel, factually-labeled database
(Supervisor: Asst. Prof. Dr. Hamdi Dibeklioğlu)
Computer Engineering Department
Contemporary multimodal deception detection is limited by the amount and quality of data available. Popular benchmarks today offer labels that are determined by psychologists. While this approach is shown to produce reliable results, unless a collective agreement is reached, interpreting deceit can be error-prone due to high interpersonal and cross-cultural variations in behavioral cues. Moreover, contemporary research disregards the possibility that deception can be localized to specific intervals of a statement in favor of treating it as a cumulative label. Thus, this study first aims to offer a novel benchmark with more than 400 unique subjects, 6 truth labels, and almost 700 data samples where samples are labeled by extensive fact-checking analysis of specialized journalists, a unique labeling strategy that minimizes personal bias. To our knowledge, this is the largest video-based deceit database available to date. Second, to our knowledge, for the first time in multimodal deceit research, a localization study will be conducted to investigate whether deceit can be studied as a regression problem as a function of time. Finally, a correspondence study with other psychological findings will be conducted to further validate the analysis presented.
DATE: 21 October 2020, Wednesday @ 13:55