Performance Analysis of Motion-Sensor Behavior for User Authentication on Smartphones

Abstract

The growing trend of using smartphones as personal computing platforms to access and store private information has stressed the demand for secure and usable authentication mechanisms. This paper investigates the feasibility and applicability of using motion-sensor behavior data for user authentication on smartphones. For each sample of the passcode, sensory data from motion sensors are analyzed to extract descriptive and intensive features for accurate and fine-grained characterization of users’ passcode-input actions. One-class learning methods are applied to the feature space for performing user authentication. Analyses are conducted using data from 48 participants with 129,621 passcode samples across various operational scenarios and different types of smartphones. Extensive experiments are included to examine the efficacy of the proposed approach, which achieves a false-rejection rate of 6.85% and a false-acceptance rate of 5.01%. Additional experiments on usability with respect to passcode length, sensitivity with respect to training sample size, scalability with respect to number of users, and flexibility with respect to screen size were provided to further explore the effectiveness and practicability. The results suggest that sensory data could provide useful authentication information, and this level of performance approaches sufficiency for two-factor authentication on smartphones. Our dataset is publicly available to facilitate future research.

Strengthen user authentication on mobile devices by using user’s touch dynamics pattern

“…Surprisingly, for one of the 8-digit PIN ("12598416"), the EER value was 4.45%, marginally worse than the 4-digit PIN ("3244"). This finding is in contrary to previous studies (Chang et al 2015;Praher and Sonntag 2016;Shen et al 2016) which have suggested that longer input strings produce a better accuracy performance.…”

“…It is not clear if the same level of accuracy performance could still be achieved when a smaller number of input samples are used. A similar case can be said for the work reported in papers (Coakley et al 2016;Roh et al 2016;Shen et al 2016).…”

“…More recently, Shen et al (2016) investigated the accuracy performance of touch dynamics on 4-digit, 5-digit, and 6-digit PINs across different operational scenarios. Unlike other related work, they used only motion features extracted from accelerometer and gyroscope sensors.…”

Convolutional Neural Networks for User Identification Based on Motion Sensors Represented as Images

“…The samples are compared based on statistical features extracted in the time domain or the frequency domain, without using machine learning. More recent studies explored the task of user identification based on machine learning models [8][9][10][11][12][13][14][15][16][17][18][19][20][21][22], attaining better results compared to statistical models such as [31]. By modeling the user identification task based on motion sensors as a classification task, various models following the standard training and evaluation pipeline used in machine learning can be tested out.…”

“…In this subsection, we present in detail the baseline methods. The first baseline method is based on a recent work by Shen et al [18], that uses handcrafted features. The second baseline method is based on another recent work by Neverova et al [17], that uses features learned by Long-Short Term Memory (LSTM) networks [49] with convolutional layers (ConvLSTM).…”