Resource-efficient federated machine unlearning via evolutionary synaptic pruning for cloud-based distributed...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Growing emphasis on user data control and privacy regulations (e.g., GDPR) necessitates effective machine unlearning (MU).
• Federated Learning (FL) presents unique challenges for unlearning due to distributed data, leading to Federated Unlearning (FU).
• Retraining models for data deletion requests in large-scale FL is computationally expensive and disruptive.

Purpose of the Study:

• To propose PRUNE-FL, a novel framework for privacy-preserving Federated Unlearning.
• To develop a method that efficiently deletes targeted data influence without full model retraining.
• To balance model performance with effective data forgetting in FL settings.

Main Methods:

• PRUNE-FL employs relevance-guided pruning and evolutionary optimization.
• A synaptic relevance scoring system identifies parameters linked to target data.
• A multi-objective problem formulation balances performance and forgetting, optimized by a genetic algorithm.

Main Results:

• PRUNE-FL demonstrates higher accuracy on the CIFAR-10 dataset in both IID and non-IID settings.
• The framework effectively removes the influence of targeted data.
• PRUNE-FL shows robustness against backdoor triggers.

Conclusions:

• PRUNE-FL offers an efficient solution for Federated Unlearning, enhancing privacy.
• The method reduces computational costs and resource usage in federated environments.
• PRUNE-FL selectively unlearns data while maintaining model performance and security.