Deep Learning Approaches for Continuous Authentication Based on Activity Patterns Using Mobile Sensing

Area of Science:

• Computer Science
• Cybersecurity
• Machine Learning

Background:

• Smartphones store sensitive personal data, making them vulnerable to security and privacy risks.
• Existing authentication methods like PINs and biometrics are susceptible to various attacks.
• Continuous authentication is needed to address evolving mobile security threats.

Purpose of the Study:

• To introduce DeepAuthen, a novel continuous authentication framework for smartphones.
• To enhance smartphone security by continuously verifying user identity through physical activity patterns.
• To evaluate the effectiveness of deep learning models in continuous user authentication.

Main Methods:

• Developed DeepAuthen, a framework utilizing smartphone accelerometer, gyroscope, and magnetometer data.
• Employed deep learning classifiers, including a proposed DeepConvLSTM network, for user authentication.
• Tested the framework on benchmark datasets: UCI-HAR, WISDM-HARB, and HMOG.

Main Results:

• Combining motion sensor data achieved high accuracy and energy efficiency ratio (EER) in binary classification.
• The DeepAuthen framework demonstrated efficacy in continuous authentication scenarios.
• Deep learning models, particularly DeepConvLSTM, proved effective for motion-based user identification.

Conclusions:

• DeepAuthen offers a robust solution for continuous smartphone user authentication.
• Physical activity patterns captured by motion sensors are valuable for enhancing mobile security.
• Deep learning techniques significantly improve the accuracy and efficiency of continuous authentication systems.