Secure deep learning for distributed data against malicious central server.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Deep learning models require secure training environments, especially with distributed data.
• Protecting central servers from malicious activities is crucial for data integrity.
• Current distributed training methods lack robust security and flexibility.

Purpose of the Study:

• To propose a novel secure system for distributed deep learning.
• To enhance security by enabling trainers to detect malicious server activities.
• To support both vertical and horizontal neural network training within a single system.

Main Methods:

• Developed a secure system connecting distributed trainers to a central parameter server.
• Implemented mechanisms for trainers to monitor and detect server-based malicious activities.
• Designed the system to accommodate both vertical and horizontal distributed training paradigms.

Main Results:

• The proposed system demonstrated the capability for distributed trainers to identify malicious server activities.
• Achieved comparable or superior area-under-the-curve scores on medical imaging datasets (MRI, X-ray) compared to existing methods.
• Successfully applied both vertical and horizontal neural network training using the developed system.

Conclusions:

• The proposed secure system enhances the safety and integrity of distributed deep learning.
• The system's flexibility in supporting both training types and its effectiveness in medical imaging analysis are significant.
• This approach offers a promising solution for secure and versatile distributed deep learning applications in healthcare.