You Are Not Acting Like Yourself: A Study on Soft Biometric Classification, Person Identification, and Mobile Device Use

In this paper, we explore the soft biometric classification of 13 demographic and behavioral attributes using phone data collected from 46 subjects. We utilize the results of this analysis to further evaluate reduced search spaces for the primary identification task by implementing a ranking formula which quantifies the performance of each attribute considering the predictability of some attributes compared to others. Results show that soft biometric classification is feasible with up to 90% accuracy; however, because people exhibit high intra-class variance, templates and queries are significantly affected in terms of how well they match, even with reduced search spaces. We analyze these findings using a combination of approaches, including an evaluation of the biometric menagerie, visualizing the distribution of the data, and observing how subjects vary in their soft biometric class across different times of the day.