SenSec: Mobile security through passive sensing

A Hybrid Deep Learning System for Real-World Mobile User Authentication Using Motion Sensors

“…After that, we integrated the previous work [ 22 , 24 , 25 , 27 , 28 , 31 , 32 , 33 , 35 , 40 , 41 , 42 , 46 ] and selected nine types of features, including RMS, standard deviation, variance, range, skewness, kurtosis, mean, average deviation, and permutation entropy. We denote the ith value of the feature vector as , which includes P features, with feature extraction of 0.2 s data in each sliding window.…”

“…De-noising ability. Most of the previous user authentication studies [ 22 , 23 , 24 , 25 , 26 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 , 42 , 47 , 48 , 49 , 50 , 51 ] considered motion sensors ideally error-free during data collection and they had never taken the noise impact of the hardware into account, which would lead to the difficulty in fitting the model and thus affecting prediction accuracy. Existing work based on large-scale real data [ 32 ] only filters the data lying on a flat surface, but in the actual environment, there are many data that are not related to authentication, such as sensor signal discontinuities and mutations, so the de-noising ability is limited.…”

“…Most of the existing feature extraction methods are based on statistics, and empirically extract a variety of time-domain and frequency-domain features, including average, standard deviation, maximum/minimum, etc. [ 22 , 24 , 25 , 27 , 28 , 31 , 32 , 33 , 35 , 37 , 38 , 40 , 41 , 42 , 46 , 47 , 51 ]; these type of features mainly rely on manual extraction and have limited ability to represent user behavior. Third, the existing sensor data modeling methods based on deep learning [ 29 , 30 , 43 , 44 , 45 , 48 , 49 , 50 ] are mostly based on the original signal data, with less consideration for signal enhancement, and lack of deeper analysis of user behavior signal data.…”

“…Although these two commonly used methods (knowledge-based authentication and static biometric-based authentication) have been in full swing for decades, there are a series of deficiencies in usability, security, and effectiveness. Therefore, in recent years, using motion sensors such as acceleration sensors, gyro sensors, and gravity sensors as data sources, a method of authentication based on the user’s dynamic behavior, has been proposed by many researchers (e.g., gait authentication [ 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 ], user’s usage behavior authentication [ 32 , 33 , 34 , 35 , 36 , 37 , 38 ], and daily life behavior authentication [ 39 , 40 , 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 , 50 , 51 ]). These methods have better user experience and stronger privacy protection capabilities, laying the foundation for the development of the third phase of mobile user authentication.…”

“…(2) In a real complex environment, the sensor data collected often lacks a trusted label. This makes supervised learning [ 22 , 24 , 25 , 27 , 28 , 32 , 33 , 35 , 37 , 38 , 40 , 41 , 42 , 47 , 48 , 49 , 50 , 51 ] difficult to use, while unsupervised learning [ 31 , 46 ] has a delay in the training phase, although a semi-supervised algorithm has been proposed in frontier work [ 32 ] to improve the accuracy, the method only uses a single classifier (SVM), and its proposed initial classifier trained with some labeled samples is a weak classifier. If the previous iteration is used continuously to classify the unlabeled data, which will cause the continuous accumulation of its own errors, then it cannot guarantee that the resulting model has high accuracy.…”