A novel gradient inversion attack framework to investigate privacy vulnerabilities during retinal image-based...

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Machine learning models for retinal image analysis show diagnostic promise.
• Federated learning (FL) enhances training data diversity but raises privacy concerns.
• Gradient inversion attacks threaten patient data privacy in FL.

Purpose of the Study:

• To evaluate the vulnerability of federated deep learning models using retinal images to gradient inversion attacks.
• To assess the impact of publicly available data on reconstructing private retinal images.
• To determine the effectiveness of current privacy countermeasures against data reconstruction.

Main Methods:

• Developed a novel framework to assess FL model vulnerability to gradient inversion attacks.
• Utilized image-to-image translation with public data to enhance reconstructed image quality.
• Evaluated reconstruction similarity against real fundus images using ResNet-18, VGG-16, and DenseNet-121 architectures.

Main Results:

• Over 92% of participants were identifiable from reconstructed retinal vessel structures across all tested models.
• Significant patient information was extractable even after implementing differential privacy.
• Reconstructed images showed high similarity to original fundus images.

Conclusions:

• Federated learning models for retinal image analysis are highly vulnerable to gradient inversion attacks.
• Existing differential privacy methods are insufficient to fully protect patient data in this context.
• Urgent development of enhanced defensive strategies is required to ensure patient privacy in federated retinal image analysis.