Federated Recommendation meets Machine Unlearning

Recommendation systems personalise online suggestions using user interactions and preferences. With rising privacy concerns and mobile device usage, federated recommender systems offer a privacy-preserving alternative by training models locally and sharing only updates. However, some cases require removing specific user contributions, a process called "unlearning," which existing methods fail to address in federated settings. We propose CFRU, a federated recommendation unlearning framework that efficiently removes target client updates without retraining. By rolling back historical updates and using sampling strategies to retain only critical ones, our method reduces storage and computational costs. We also estimate and compensate for bias introduced by update removal, ensuring minimal performance loss. Experiments on real-world datasets, including adversarial scenarios, show that CFRU achieves 99.3% of full retraining quality while being up to 1000 times faster.