Computationally Efficient Federated Unlearning

Federated learning (FL) must support a client's right to be forgotten (a.k.a Federated Unlearning), enabling clients to have their data influence removed from a trained model. However, existing Federated Unlearning (FU) algorithms often incur high overhead, especially those based on influence functions that require computing secondorder Hessian matrices. We propose a novel algorithm to efficient FU that eliminates the need for any Hessian computation or inversion. Instead, our algorithm uses only first-order gradients to approximate each training point's influence via a first-order Taylor expansion. By avoiding expensive Hessian computations, we reduce the computational complexity of unlearning from cubic to roughly quadratic time (O (n(3)) to similar to O (n(2)) in the number of model parameters). Our algorithm operates locally on the client's data, preserving privacy by not involving other clients or raw data exchange.