BlindU: Blind Machine Unlearning Without Revealing Erasing Data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Machine unlearning aims to remove data's influence from trained models for privacy.
• Current unlearning methods often require uploading data to the server, posing privacy risks, especially in federated learning (FL).

Purpose of the Study:

• To develop a machine unlearning method that operates without the server accessing the original data to be erased.
• To ensure privacy-preserving unlearning in scenarios like FL where direct data access is prohibited.

Main Methods:

• Proposing Blind Unlearning (BlindU), which utilizes compressed, privacy-preserving representations for unlearning.
• Integrating an information bottleneck (IB) mechanism in FL for local generation of compressed representations.
• Employing dedicated unlearning modules for IB-based models and multiple gradient descent for balancing forgetting and utility.
• Introducing a noise-free differential privacy (DP) masking method for enhanced privacy before compression.

Main Results:

• BlindU effectively performs unlearning using compressed representations, safeguarding raw data from the server.
• The method demonstrates superior privacy protection compared to existing privacy-preserving unlearning benchmarks.
• Experimental results validate the effectiveness of BlindU in both privacy and unlearning performance.

Conclusions:

• Blind Unlearning (BlindU) offers a feasible solution for private machine unlearning in sensitive environments like federated learning.
• The approach successfully decouples the unlearning process from direct data access, enhancing user privacy.
• BlindU represents a significant advancement in privacy-preserving machine learning techniques.