FedActUnlearn proposes a novel activation perturbation aggregation framework for scalable machine unlearning in federated environments. The method encodes client contributions through lightweight perturbation signatures, enabling efficient removal without retraining or extensive storage. It achieves (ϵ, δ)-certified unlearning with theoretical guarantees on approximation quality. Experimental evaluation on FEMNIST, CIFAR-10, and Shakespeare demonstrates comparable accuracy to exact retraining with 95× reduction in unlearning latency and 50× lower storage requirements.
Key findings
FedActUnlearn achieves (ϵ, δ)-certified unlearning under standard convexity assumptions with explicit bounds on approximation error relative to exact retraining.
The method reduces unlearning latency by up to 95× compared to retraining-based approaches while maintaining comparable model accuracy.
Storage overhead is reduced by 50× versus historical caching methods, requiring only O(d) space per client independent of training rounds.
Client removal is accomplished in O(1) communication rounds compared to O(T) rounds required by retraining methods.
The framework is compatible with secure aggregation protocols and immediately deployable in production federated learning systems.
Limitations & open questions
Theoretical guarantees assume convexity conditions that may not strictly hold for deep neural networks.
Experimental validation is limited to benchmark datasets (FEMNIST, CIFAR-10, Shakespeare) and may not reflect all real-world federated deployment scenarios.