Adaptive feature unlearning for trustworthy medical imaging privacy

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Data Privacy

Background:

• Deep learning in medical imaging risks patient privacy due to data memorization.
• Current machine unlearning methods have limitations in feature-level forgetting, real-time application, and verification.

Purpose of the Study:

• To introduce AdaptForget, a domain-adaptive, feature-level unlearning framework for privacy-preserving medical image analysis.
• To address limitations of existing machine unlearning techniques in medical contexts.

Main Methods:

• AdaptForget utilizes out-of-distribution (OOD) guidance for feature disentanglement.
• A novel OOD-driven feature-output disentanglement loss prevents feature collapse.
• Single-entry forgetting is formalized for immediate data erasure.
• Isolation verification distance is proposed as a metric for auditing feature-level forgetting.

Main Results:

• AdaptForget achieves state-of-the-art privacy protection in medical imaging.
• The framework effectively preserves model utility after unlearning.
• Experiments across diverse medical imaging and healthcare datasets validate the approach.

Conclusions:

• AdaptForget offers a robust solution for privacy-preserving medical image analysis.
• The framework enables verifiable, feature-level machine unlearning.
• Timely revocation of individual patient data is feasible with AdaptForget.