Resource-efficient Federated Machine Unlearning Via Evolutionary Synaptic Pruning for Cloud-Based Distributed Learning Systems

With an increasing emphasis on user data control and privacy regulations, such as the General Data Protection Regulation (GDPR), machine unlearning (MU) has emerged as a crucial mechanism for managing data in AI systems. MU enables models to remove the influence of specific user data upon request. This problem becomes more complex in collaborative settings such as Federated Learning (FL), where data remains distributed across multiple clients, giving rise to Federated Unlearning (FU). In large-scale deployments, particularly those supported by cloud infrastructure, retraining models to satisfy data deletion requests can be computationally expensive, energy-intensive, and disruptive to ongoing services. This highlights the importance of having effective methods to unlearn outdated practices that can hinder the growth of a system and compromise data privacy awareness. We propose PRUNE-FL (Privacy-preserving Retention-focused Unlearning with Neuro-Evolution in Federated Learning), a framework that uses relevance-guided pruning and evolutionary optimization to delete the influence of the targeted data. PRUNE-FL is different from methods that rely on retraining or coarse parameter updates because it focuses on finding and changing the parameters that are most closely related to the data that needs to be forgotten. The approach is based on a synaptic relevance scoring system to figure out how each model parameter relates to certain client or class-level data. This makes it easier to find parameters that are connected to the target data. Then, the unlearning task is set up as a multi-objective problem to find a balance between the overall performance and the forgetting of unnecessary information. Finally, a genetic algorithm is used to implement an evolutionary pruning strategy. It seeks the optimal pruning settings that operate within the constraints of federated learning. Thus, PRUNE-FL helps in unlearning specific data without having to retrain the whole system. Tests on the CIFAR-10 dataset in both IID (Independent and Identically Distributed) and non-IID settings show that PRUNE-FL has higher accuracy. It also effectively removes the influence of the targeted data. The results also show it to be strong against bad patterns like backdoor triggers. Overall, PRUNE-FL enhances privacy by selectively unlearning the data and using fewer resources in federated environments.