Bag-of-Lies: A Multimodal Dataset for Deception Detection

摘要

Deception detection is a pervasive issue in security. It has been widely studied using traditional modalities, such as video, audio and transcripts; however, there has been a lack of investigation in using modalities such as EEG and Gaze data due to the scarcity of a publicly available dataset. In this paper, a new multimodal dataset is presented, which provides data for deception detection by the aid of various modalities, such as video, audio, EEG and gaze data. The dataset explores the cognitive aspect of deception and combines it with vision. The presented dataset is collected in a realistic scenario and has 35 unique subjects providing 325 annotated data points with an even distribution of truth (163) and lie (162). The benefits provided by incorporating multiple modalities for fusion on the proposed dataset is also investigated. It is our assertion that the availability of this dataset will facilitate the development of better deception detection algorithms which are more relevant to real world scenarios.