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

Area of Science:

• Computer Science
• Cybersecurity
• Machine Learning

Background:

• Mobile devices are ubiquitous, raising concerns about private data security and user authentication.
• Existing motion sensor-based authentication methods suffer from poor de-noising, limited availability, and inadequate feature extraction.
• There is a need for robust and user-friendly mobile authentication solutions.

Purpose of the Study:

• To propose a hybrid deep learning system for enhanced mobile user authentication.
• To address the limitations of current motion sensor-based authentication techniques.
• To improve the accuracy and reliability of mobile device security.

Main Methods:

• Implemented a variational mode decomposition (VMD) based de-noising method to improve sensor data quality.
• Utilized semi-supervised collaborative training (Tri-Training) to handle real-world data mislabeling.
• Developed a hybrid model combining convolutional neural networks (CNN) and support vector machines (SVM) for feature extraction and training.

Main Results:

• The proposed system achieved a high authentication accuracy of 95.01% on large-scale, real-world data.
• The VMD method enhanced singular values and expanded the feature extraction range.
• The hybrid CNN-SVM model demonstrated effective feature extraction and training capabilities.

Conclusions:

• The hybrid deep learning system offers a superior solution for complex, real-world mobile authentication.
• The integration of VMD and Tri-Training significantly improves authentication performance.
• This approach provides a more secure and available method for protecting private data on mobile devices.