Application of Machine Learning in Deception Detection

Abstract

The issue of security is a continuous struggle for all. To address this struggle, it is pertinent to reliably detect deception. To reliably detect deception is a knotty task as no ideal technique has been found for the detection. According to literature, past researches focused on single cue, it was observed that combining cues will significantly be a good indicator of deception that using a single cue. Since no single verbal or non-verbal cue is able to detect deception successfully the research proposes to combine verbal and non-verbal cues for the detection. Therefore, this research aims to develop a neurofuzzy model for classifying extracted verbal and nonverbal features as deceptive or truthful. The proposed system extracted desired features from the dataset of Perez-Rosas. The verbal cues include the voice pitch, jitters, pauses, and speechrate. The PRAAT was used in extracting all the verbal cues. The nonverbal features were extracted using the Active Shape Model (ASM) and the classification Model was designed using Neurofuzzy technique. The work was implemented in 2015a MatLab. The developed model was compared with Support Vector Machine (SVM), K-Nearest Neighbour (KNN) and Decision Tree. Neurofuzzy recorded the best performance with the Nonverbal dataset (percentage score of 97.1%), KNN performed well with the Verbal dataset (percentage score of 90.9%) while Decision Tree performed best with the VerbNon dataset (percentage score of 97.6%). From the comparative analysis it was discovered that Neurofuzzy model work well on Nonverbal dataset to detect deception. The result obtained using only verbal cue was 84.3% while that of nonverbal cue was 97.1% but on VerbNon it yielded 92.5% which is far better than the chance level of 50%.