Towards a non-intrusive behavioral identification using hand movements and deep neural networks

Abstract

Studies of hand behavior-based implicit authentication have applied various machine learning classifiers to analyze the behaviors. Yet, there are a few studies about deep learning-based authentication. The purpose of this research is to explore a user identification method based on ResNet classifier and hand behavioral patterns, and to seek an opportunity of expansion to authentication. The ResNet classifier is applied in three tasks to prove its availability on classification and identification. From two independent experiments, either simple hand behavior or complex hand behavior are measured and used in one of the three tasks. An IMU embedded smart clips called MANOVIVO collects accelerations and Euler angles. The data expands to a 29-feature set consisted of accelerations, velocities, Euler angles, and joint angles through data preprocessing. Through the explained experiments and tasks, ResNet-101, which contains 101 convolution layers, shows 7.45 % of EER on identification task along with 738 of simple hand behavior data, and 11.11 % of EER on identification task along with 148 of complex hand behavior data. The hand behavior-based identification with a large dataset is modifiable to implicit authentication which may apply to the smartphone and its mobile applications’ security process without any biases on race and gender.