Development of a Deep-Learning-Based Computerized Scoring Algorithm

TL;DR

This paper introduces a deep-learning-based system to improve the accuracy of polygraph tests by reducing human bias and handling complex biological signals.

Contribution

A novel deep neural network-based computerized scoring system for polygraphs is developed to address limitations of traditional linear models.

Findings

The deep-learning model achieved high recall, precision, and F1 scores of 0.9681, 0.9700, and 0.9683, respectively.

The system outperformed conventional linear classifier-based polygraph scoring methods.

The model effectively handles nonlinear bio-signals, reducing examiner bias.

Abstract

During polygraph tests, the examiner evaluates physiological responses recorded on a chart to identify deception. Generally, this evaluation involves a numerical scoring system. However, biases related to politics, region, and religion, as well as personal factors such as fatigue and stress, can lead to inaccuracies in the examiner’s judgment. To solve these problems, computerized scoring systems (CSSs) that automatically analyze charts have been introduced, aiming to reduce human error. Conventional CSS models, which rely on linear classifiers, struggle with the nonlinear nature of biological signals, resulting in poor performance. Therefore, it is crucial to incorporate deep learning structures such as deep neural networks, which account for the nonlinearity of bio-signals, to enhance effectiveness of CSSs. This paper introduces a Korean computerized scoring system that leverages a deep neural network, which was developed to mitigate the subjective bias of polygraph examiners and to obtain high-accuracy results by considering the nonlinearity of bio-signals. The performance of the developed algorithm was evaluated, demonstrating recall, precision, and F1 scores of 0.9681 ± 0.0314, 0.9700 ± 0.0321, and 0.9683 ± 0.0171, respectively. These results suggested a significant improvement in CSS performance over conventional systems that depend on linear classifiers.

Figures