1571302378 Evaluating Federated Unlearning for Classification

Evaluating Federated Unlearning for Classification: A Comparative Study and Trade-off Analysis Paper Title: Evaluating Federated Unlearning for Classification: A Comparative Study and Trade-off Analysis Authors: Nina Abeyratne, Nipuna Senanayake Institution: School of Computing, Informatics Institute of Technology (IIT), Colombo, Sri Lanka Conference: 2026 IEEE International Conference on Artificial Intelligence in Engineering and Technology (IICAIET 2026) This presentation covers a unified comparative evaluation of centralized and federated machine unlearning methods, addressing the growing need for privacy-compliant AI systems under regulations such as GDPR and CCPA. The study evaluates methods including Gradient Ascent, First Epoch Reversal, FRAMU, NoT, and FedEraser across three datasets — MNIST, CIFAR-10, and ChestMNIST; under IID and non-IID data distributions. Performance is assessed across five dimensions: utility, forgetting effectiveness, privacy (via Membership Inference Attack), computational efficiency, and communication cost. Key findings include the identification of an efficiency-effectiveness tension in centralized unlearning, FRAMU as the most balanced federated method with 80% communication reduction, and a scenario-based deployment guide for privacy-critical, resource-constrained, and large-scale federated systems. Presented at IICAIET 2026 - IEEE International Conference on Artificial Intelligence in Engineering and Technology.