Federated Unlearning: A Survey on Methods, Design Guidelines, and Evaluation Metrics

Area of Science:

• Machine Learning
• Data Privacy
• Cybersecurity

Background:

• Federated learning (FL) trains models collaboratively while keeping data local, enhancing privacy and GDPR compliance.
• Existing FL lacks clear methods for data removal (right to be forgotten) and defense against malicious client backdoors.
• Removing specific data contributions without full model retraining is a significant challenge.

Purpose of the Study:

• To address the need for efficient federated unlearning (FU) algorithms.
• To enable the removal of specific data contributions and malicious updates from FL models.
• To provide practical guidelines and a taxonomy for FU schemes.

Main Methods:

• Literature review of state-of-the-art federated unlearning contributions.
• Analysis of metrics for evaluating unlearning effectiveness in FL.
• Categorization of FU methods under a novel taxonomy.

Main Results:

• Identified the necessity for novel federated unlearning algorithms.
• Provided background concepts, empirical evidence, and practical guidelines for FU.
• Detailed analysis of FU evaluation metrics and a comprehensive literature review.

Conclusions:

• Federated unlearning is crucial for privacy-preserving AI, enabling data removal and backdoor defense.
• Efficient FU schemes are essential to maintain model integrity and knowledge without complete retraining.
• Open technical challenges and future research directions in federated unlearning were outlined.