DiffNet: An EEG-based Lie Detection framework by Integrating Stockwell Transform with CNN-DTLSTM using ERP P300 signals

Lie detection is essential as it can uncover deceit across various domains, including security, psychological research, and law enforcement. However, the accuracy and effectiveness of traditional strategies are constrained. This study addresses this issue by proposing a novel approach for lie detection that utilizes P300 signals and deep learning techniques. Stimuli associated with deceptive or honest responses produce P300 signals that serve as crucial physiological markers. The analysis of intricate P300 patterns is improved by utilizing deep learning models, thereby increasing the accuracy of detection. In this study, electroencephalography (EEG) signals from 42 individuals were recorded as they underwent a concealed information test (CIT). Subsequently, Stockwell Transform (S-Transform) method was employed to elicit key features from the preprocessed EEG data. These features were then fed into a CNN and a tanh LSTM (TLSTM) structure incorporating an input differential characteristic (DTLSTM) module, forming the CNN-DTLSTM architecture, named DiffNet, for discerning between truthful and deceptive samples. The resultant accuracy obtained from the proposed lie detection system is 98.25%.