Federated Unlearning (FU) aims to delete specific training data from an ML model trained using Federated Learning (FL). However, existing FU methods suffer from inefficiencies due to the high costs associated with gradient recomputation and storage. This paper presents QuickDrop, an original and efficient FU approach designed to overcome these limitations. During model training, each client uses QuickDrop to generate a compact synthetic dataset, serving as a compressed representation of the gradient information utilized during training. This synthetic dataset facilitates fast gradient approximation, allowing rapid downstream unlearning at minimal storage cost. To unlearn some knowledge from the trained model, QuickDrop clients execute stochastic gradient ascent with samples from the synthetic datasets instead of the training dataset. The tiny volume of synthetic data significantly reduces computational overhead compared to conventional FU methods. Evaluations with three standard datasets and five baselines show that, with comparable accuracy guarantees, QuickDrop reduces the unlearning duration by 463× compared to retraining the model from scratch and 65 -- 218× compared to FU baselines. QuickDrop supports both class- and client-level unlearning, multiple unlearning requests, and relearning of previously erased data.

Federated Learning (FL) has gained considerable attention recently, as it allows clients to cooperatively train a global machine learning model without sharing raw data. However, its performance can be compromised due to the high heterogeneity in clients' local data distributions, commonly known as Non-IID (non-independent and identically distributed). Moreover, collaboration among highly dissimilar clients exacerbates this performance degradation. Personalized FL seeks to mitigate this by enabling clients to collaborate primarily with others who have similar data characteristics, thereby producing personalized models. We noticed that existing methods for assessing model similarity often do not capture the genuine relevance of client domains. In response, our paper enhances personalized client collaboration in FL by introducing a metric for domain relevance between clients. Specifically, to facilitate optimal coalition formation, we measure the marginal contributions of client models using coalition game theory, providing a more accurate representation of potential client domain relevance within the FL privacy-preserving framework. Based on this metric, we then adjust each client's coalition membership and implement a personalized FL aggregation algorithm that is robust to Non-IID data domain. We provide a theoretical analysis of the algorithm's convergence and generalization capabilities. Our extensive evaluations on multiple datasets, including MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100, and under varying Non-IID data distributions (Pathological and Dirichlet), demonstrate that our personalized collaboration approach consistently outperforms contemporary benchmarks in terms of accuracy for individual clients.