An 'object-use fingerprint': the use of electronic sensors for human identification

Abstract

We discovered an error in the data analysis after the publication of this paper. The corrected results are presented in the attachment. With the consent of the Ubicomp program chairs, these corrections were distributed and presented at the conference. Abstract. We describe an experiment in using sensor-based data to identify individuals as they perform a simple activity of daily living (making coffee). The goal is to determine whether people have regular and recognizable patterns of interaction with objects as they perform such activities. We describe the use of a machine-learning algorithm to induce decision-trees that classify interaction patterns according to the subject who exhibited them; we consider which features of the sensor data have the most effect on classification accuracy; and we consider ways of reducing the computational complexity introduced by the most important feature type. Although our experiment is preliminary, the results are encouraging: we are able to do identification with an overall accuracy rate of 97%, including correctly recognizing each individual in at least 9 of 10 trials. 1