Margin-aware prototype learning for client withdrawal in federated unlearning.

Area of Science:

• Machine Learning
• Federated Learning
• Data Privacy

Background:

• Federated learning enables collaborative model training without sharing raw data.
• Client withdrawal requires removing a participant's influence while maintaining model utility for others.
• Current methods for client withdrawal involve trade-offs between efficiency, memory usage, and complete influence removal.

Purpose of the Study:

• To develop an efficient and effective framework for federated client withdrawal.
• To address the limitations of existing methods that rely on historical data or are computationally expensive.
• To ensure the utility of the collaboratively trained model is preserved for remaining participants.

Main Methods:

• Introduced Margin-Aware Prototype Learning (MAPLE), a novel framework for federated client withdrawal.
• Employed Margin-aware Label Reassignment (MLR) to locally perturb labels on withdrawing client data, focusing on low-confidence samples.
• Utilized Prototype-driven Constraints (ProCons) with class-wise prototypes from remaining clients to preserve global knowledge via contrastive learning.

Main Results:

• MAPLE achieves unlearning quality comparable to complete retraining.
• The framework demonstrates significant efficiency, being orders of magnitude faster than retraining.
• MAPLE requires negligible memory overhead, outperforming state-of-the-art approaches.
• Experimental results show MAPLE effectively removes client influence while preserving model utility.

Conclusions:

• MAPLE offers an efficient and effective solution for federated client withdrawal.
• The proposed method overcomes the efficiency-utility trade-off inherent in existing approaches.
• MAPLE provides a practical alternative to computationally prohibitive retraining methods for client unlearning.