Overview

Federated unlearning studies how to remove the influence of clients or individual records from models trained across distributed data, without simply rebuilding the model from scratch. Approaches include retraining and replay, influence- or gradient-based updates, exact deletion from additive statistics, and lightweight methods such as pruning, filters, distillation, or proxy-data calibration. Methods are tested in settings including non-IID image and medical data, graphs, personalized models, multimodal systems, and federated language models, often against retraining as a reference. A central tension is making deletion fast and preserving utility while ensuring forgotten information is truly gone: output accuracy alone may miss residual representations or membership signals, and server-side summaries can themselves leak data.

Latest papers

31 papers5 papers in the last four weeks, up 67% on the four weeks before. 0.0% of all new papers.

Mar 13, 2026

Exact Federated Continual Unlearning for Ridge Heads on Frozen Foundation Models

Foundation models are commonly deployed as frozen feature extractors with a small trainable head to adapt to private, user-generated data in federated settings. The ``right to be forgotten'' requires removing the influence of specific samples or users from the trained model on demand. Existing federated unlearning methods target general deep models and rely on approximate reconstruction or selective retraining, making exactness costly or elusive. We study this problem in a practically relevant but under-explored regime: a frozen foundation model with a ridge-regression head. The exact optimum depends on the data only through two additive sufficient statistics, which we turn into a communication protocol supporting an arbitrary stream of add and delete requests via fixed-size messages. The server maintains a head that is, in exact arithmetic, pointwise identical to centralized retraining after every request. We provide deterministic retrain-equivalence guarantees, order and partition invariance, two server-side variants, and a Bayesian certificate of zero KL divergence. Experiments on four benchmarks confirm the guarantees: both variants match centralized ridge retraining to within 10−910^{-9} relative Frobenius error and complete each request at orders-of-magnitude lower cost than federated retraining baselines.