Top-1 compression suffices for federated unlearning with the help of adaptive error feedback (🎉 Accepted by ICASSP 2026, unpublished)
See EFFACE code overview for more details.
More Results on SVHN and Fashion-MNIST datasets, please check results.pdf