Designing an Effective Deception Classification System Using Hybrid TLBO-DE-Based Optimization and Residual Temporal Convolutional Networks

Deception is a well-known term that refers to acting in a way that causes another person to believe something that is not true. Deception is a rather common occurrence in everyday life and frequently becomes a major national security concern. When investigating crimes, it’s very crucial to be able to detect dishonesty accurately. As a remedy to this problem, deception detection has recently attracted a lot of attention. In this paper, we sought to construct a deception detection system using electroencephalography (EEG) data gathered from a Concealed Information Test (CIT). To improve the system’s performance, we proposed a hybrid Teaching Learning Based Optimization-Differential Evolution (TLBO-DE) with a novel fitness function to select the best channels. Further, the DE mutant and difference vectors are modified for faster convergence while balancing the exploration and exploitation. In addition, a Residual Temporal Convolution Network (RTCN) is used to reduce model complexity without compromising the performance. The proposed approach is experimented on CIT dataset. The TLBO-DE hybrid channel selection achieved better results in terms of number of channels and accuracy-related measures when compared with TLBO, DE-based channel selection methods. Further, the proposed channel selection model with RTCN is compared with other deep learning models like LSTM, RNN and GRUs. The results proved that it not only improves accuracy, sensitivity, F1-score, and specificity but also reduced model complexity. The proposed system’s performance resulted in 96.30% accuracy and is outperformed compared to State-of-The-Art (SOTA) models.