Forgettable Federated Linear Learning With Certified Data Unlearning

Summary

Researchers introduce forgettable federated linear learning (F^2L^2), a federated training strategy for deep neural networks that approximates them linearly using pretrained models. They also present FedRemoval, an unlearning method that lets the server remove a target client without client communication or extra storage. Experiments on small to large datasets, using convolutional networks and foundation models, show the approach balances accuracy with successful unlearning.