FedDRU: Depth-Aware Residual Client Unlearning for Heterogeneous Federated Learning

Abstract

Client-level federated unlearning seeks to update an already trained global model so that the influence of a specified client is weakened, while the model remains effective for the remaining clients. Existing methods are mostly designed for homogeneous model settings and often rely on retraining, historical updates, or direct gradient reversal, making them less suitable for clients with heterogeneous computational capacities. This paper studies client-level unlearning in heterogeneous multi-exit federated learning, where clients participate with different model depths. We propose FedDRU, a depth-aware residual unlearning method that uses retained-client directions to represent shared knowledge, decomposes the target-client update into shared and residual components, and reverses only the residual contribution with a lightweight confidence softening constraint. Experiments on CIFAR-10 and CIFAR-100 with three-exit ResNet and ViT models show that FedDRU effectively suppresses target-client influence while maintaining competitive retained-client accuracy, achieving a stable trade-off between forgetting effectiveness and retained-client utility.