Sensitivity-Aware Differential Privacy for Federated Medical Imaging

Area of Science:

• Artificial Intelligence in Healthcare
• Privacy-Preserving Machine Learning
• Medical Imaging Analysis

Background:

• Federated learning (FL) facilitates collaborative AI model training across institutions without raw data sharing, ideal for smart healthcare.
• Gradient inversion attacks (GIAs) pose a privacy risk in FL, as private information can be inferred from shared gradients.
• Traditional differential privacy (DP) offers uniform protection, often leading to suboptimal performance and increased privacy risks for sensitive data.

Purpose of the Study:

• To introduce a novel privacy notion, sensitivity-aware differential privacy, to enhance the balance between model performance and privacy protection in FL.
• To address the limitations of uniform privacy protection in traditional DP methods for healthcare applications.

Main Methods:

• Proposed a sensitivity-aware differential privacy framework where privacy protection is adjusted based on objective measurements of data sample sensitivity.
• Developed a defense mechanism that dynamically modifies privacy protection levels in response to varying privacy leakage risks from GIAs.
• Extended the proposed method to effectively handle multi-attack scenarios.

Main Results:

• Demonstrated the efficacy of the sensitivity-aware approach through extensive experiments on real-world medical imaging datasets.
• Achieved an average performance improvement of 13.5% compared to state-of-the-art methods under equivalent privacy risk.
• Showcased improved model performance and enhanced privacy guarantees by tailoring protection to data sensitivity.

Conclusions:

• Sensitivity-aware differential privacy offers a more effective approach to privacy protection in federated learning for healthcare.
• The proposed method significantly improves model performance while maintaining robust privacy guarantees against sophisticated attacks.
• This approach represents a significant advancement in secure and efficient collaborative machine learning for sensitive medical data.