Learning and Unlearning Your Data in Federated Settings
summary
Federated unlearning requires methods that approximate full retraining because exact machine unlearning is impractical at scale; the proposal analyzes multiple algorithms, compares them by consistency, accuracy, provable privacy and efficiency, and flags that differential privacy is necessary but not sufficient (even epsilon=18 degrades accuracy). It examines the certified removal metric, highlights non‑IID data, client churn and who actually performs removal, contrasts passive vs active unlearning approaches, and notes missing proofs-of-removal and vulnerability to model inference attacks.
AI-written summary · about
chapters (5)
- 0:00Machine unlearning intro
- 2:42Aims of unlearning
- 7:28Unlearning algorithm requirements
- 8:30Federated unlearning challenges
- 13:22Open problems in federated unlearning
related talks
30mUnforgettable, that's what you are: Evaluating Machine Unlearning and Forgetting
29mBalancing Thin Line Between Data Intelligence & Privacy
23mHow can a learnt ML model unlearn something: Framework for "Machine Unlearning"
09mInterview with Blaise Aguéra y Arcas
18mRectifying Privacy and Efficacy Measurements in Machine Unlearning...
34mDecentralized Machine Learning
41mFederated Learning: Machine Learning on Decentralized Data
31mFederated Learning: So Close and Yet So Out of Reach