Multimodal detection of concealed information using Genetic-SVM classifier with strict validation structure

Abstract

The main idea of multimodal detection of concealed information is based on a comparison between the physiological responses of a subject to the probe and irrelevant items. The purpose of the present study was to extend the previous study by using a new machine learning structure to discriminate guilty and innocent subjects. The new machine learning structure was designed based on utilizing genetic support vector machine (GSVM) as a classifier. Evaluation of GSVM was performed through a strict validation structure to prevent from optimistic result. But this method entailed the problem of finding several optimal parameters in one optimization problem. To solve this problem, an innovative approach based on “the most frequent result” was introduced, and therefore the general optimal parameters in each feature subset were obtained. The classification accuracy of GSVM based on ERP feature was approximately 90%. The best accuracy among autonomic features and their combinations was around 95.45% and among combinations of ERP and autonomic features was around 93%. In comparison with previously used methods on the same data, GSVM had higher accuracy in all of the feature subsets. The new presented machine learning structure made experimental conditions closer to real conditions by using strict validation and introducing general optimal parameter. On the other hand, the higher accuracy of GSVM shows the better SVM's ability to find complex patterns and solve classification problems as compared with the other methods. Moreover, the crucial role of Genetic algorithms to find the optimal parameters shouldn't be ignored.

Keywords

• Genetic-SVM
• Strict validation structure
• Concealed information test (CIT)
• Event-related potential (ERP)
• Autonomic response