Evaluating Federated Unlearning for Classification

BSc (Hons) Artificial Intelligence and Data Science | Individual Research Project

Nina Abeyratne

Machine unlearning enables trained models to selectively forget specific data, supporting privacy regulations such as the General Data Protection Regulation (GDPR). Although centralized unlearning has been widely explored, federated environments introduce additional challenges due to decentralized data ownership, communication overhead, and heterogeneous client distributions. This study presents a unified comparative evaluation of centralized and federated unlearning methods across multiple benchmark settings, including MNIST, CIFAR-10, and ChestMNIST, under varied deletion scales and data distributions. The evaluated methods are compared using retained model utility, forgetting effectiveness, privacy behavior under membership inference attacks (MIAs), computational cost, and communication efficiency. The results confirm that exact retraining provides the strongest forgetting reference but remains computationally impractical for repeated deployment. More importantly, the comparative analysis reveals clear trade-offs among privacy preservation, retained utility, execution cost, and communication overhead, showing that no approximate unlearning strategy consistently dominates across all conditions. Among the evaluated federated approaches, FRAMU demonstrates the most stable overall balance, while lightweight heuristic methods such as NoT become increasingly unreliable under heterogeneous non- IID settings. Similar comparative trends are further observed in the ChestMNIST medical imaging benchmark, supporting the practical relevance of these findings beyond low-complexity datasets. Overall, the study provides a multidimensional guidance framework for selecting suitable unlearning strategies according to deployment priorities.