Developing virtual fingerprint: identifying trustworthy interaction behaviors of web users

Abstract Trustworthy interaction refers to the interaction process between human and network information systems that is believable and credible. In recent years, network security events frequently happened caused by human-web interaction lacking trustworthiness, which brings about some security risks such as private information leakage, resource abuse, property or reputation loss, etc. It is a significant network security concern threatening people. The problem is typically approached from the viewpoint of identifying the trustworthiness of the credentials submitted by interactors. There are little few studies that have addressed it in terms of the trustworthiness of the interaction behavior by interactors. We propose that the interaction behavior pattern of interactors on the website is a virtual fingerprint (VF) to discriminate the trustworthiness of the identity based on relevant psychological and behavioral theories. Using 481,838 weblog data from a travel service website, we demonstrate the effectiveness of the proposed VF. First, the individual differences in interaction behavior of interactors on the website are characterized by numerical analysis and statistical inference. Then, 15 interaction behavior features are constructed from weblogs to create a VF, which can discriminate the trustworthiness of the identity by identifying the interaction behaviors of the interactors on the website. Of these, a machine learning algorithm (i.e., the eXtreme gradient boosting) is applied to build the identification model for interaction behaviors. The results show that the VF proposed in this study has better discrimination of trustworthiness compared to those approaches that use other kinds of behavior features. And, the eXtreme gradient boosting used performs more efficiently than the algorithm of Decision Tree and Random Forest. Specifically, the errors of discriminating both untrustworthy and trustworthy identities are as low as 0.19% and 3.30%, respectively. Also, the accuracy of identifying interaction behaviors is as good as 99.63%.