A Method of Information Protection for Collaborative Deep Learning under GAN Model Attack.

Area of Science:

• Medical image analysis
• Deep learning applications
• Data privacy in AI

Background:

• Deep learning (DL) offers high accuracy in medical image classification and biological applications.
• Collaborative DL training in medicine faces significant information leakage risks, particularly from generative adversarial network (GAN) attacks.
• Existing privacy protection methods are insufficient against sophisticated attacks in the medical domain.

Purpose of the Study:

• To propose a novel privacy protection method for collaborative deep learning in the medical field.
• To enhance the stability and security of deep learning models during collaborative training.
• To mitigate information leakage risks associated with generative adversarial network (GAN) attacks.

Main Methods:

• Implementation of a deep convolutional generative adversarial network (DCGAN) framework.
• Utilizing encrypted transmission for deep network parameter sharing.
• Incorporating a detection mechanism for generative adversarial network (GAN) attacks within the network.
• Dynamically adjusting training parameters to invalidate adversarial attacks.

Main Results:

• The proposed DCGAN-based method effectively protects sensitive information during collaborative deep learning.
• The method enhances the overall stability of the training process against GAN-based attacks.
• Information leakage risks are significantly reduced, ensuring data confidentiality in medical applications.

Conclusions:

• The developed privacy protection method offers a robust solution for secure collaborative deep learning in medicine.
• Encrypted transmission and adaptive parameter adjustment are key to defending against GAN attacks.
• This approach is crucial for maintaining patient data privacy and trust in AI-driven healthcare.