Video Based Deception Detection Using Deep Recurrent Convolutional Neural Network

Automatic deception detection has been extensively studied considering their applicability in various real-life applications. Since humans will express the deception through non-verbal behavior that can be recorded in a non-intrusive manner, the deception detection from video using automatic techniques can be devised. In this paper, we present a novel technique for the video-based deception technique using Deep Recurrent Convolutional Neural Network. The proposed method uses the sequential input that can capture the spatiotemporal information to capture the non-verbal behavior from the video. The deep features are extracted from the sequence of frames using a pre-trained GoogleNet CNN. To effectively learn the extended sequence, the bi-directional LSTMs are connected to the GoogleNet and can be jointly trained to learn the perceptual representation. Extensive experiments are carried out on a publicly available dataset [5] with 121 deceptive and truthful video clips reflecting a real-life scenario. Obtained results demonstrate the outstanding performance of the proposed method when compared with the four different state-of-the-art techniques.