IncreAuth: Incremental-Learning-Based Behavioral Biometric Authentication on Smartphones

Zhihao Shen,Shun Li,Xi Zhao,Jianhua Zou

DOI: https://doi.org/10.1109/jiot.2023.3289935

IF: 10.6

2023-01-01

IEEE Internet of Things Journal

Abstract:Touch behavior biometric has been widely studied for continuous authentication on mobile devices, which provides a more secure authentication in an implicit process. However, the existing touch behavior biometric-based authentication systems suffer from two issues. First, the existing touch behavior representation methods are hard to characterize touch operations under complex usage context. Second, the authentication accuracy of existing authentication models is inclined to degrade over time in a long-term real-life usage scenario due to change in data distribution caused by varying touch behavior. Toward this end, in this article, we develop IncreAuth, an incremental learning-based continuous authentication framework, which allows to provide effective stable authentication performance in the long-term smartphone usage scenario. Specifically, we first propose a novel context-aware feature set to characterize touch behavior patterns in complex usage context. Then, we develop an authentication model GBDTNN, which integrates the advantages of a gradient boosting decision tree model for processing our high-dimensional feature set and neural network model for efficient online updating. A behavior drift-based online updating mechanism is also designed to learn both long-term and short-term touch behavior patterns. To evaluate our framework, we construct a large-scale smartphone usage data set over two months collected from the unconstrained environment. Extensive experiments demonstrate that IncreAuth achieves the state-of-the-art and stable authentication accuracy over time and low system overheads.