Federated Unlearning via Distilled Data

Federated Learning (FL) has emerged as a privacypreserving paradigm that enables collaborative model training between distributed client devices without exchanging raw data. However, while FL mitigates direct data exposure, the influence of client data remains encoded in the trained model, raising concerns in scenarios where data deletion is legally or ethically required. Federated Unlearning (FU) aims to address this gap. Yet first existing FU solutions often rely on strong assumptions, such as maintaining complete histories of model updates or leveraging publicly available datasets that resemble private client data, or propose solutions which are not selective and coarsely reinitialize part of the learned model (with the associated non-negligible overhead). This work introduces a novel unlearning method based on distilled synthetic data, which clients can generate and transmit to the server; these compact and distilled samples are unrecognizable from the original data, but preserve most of their training influence, thus enabling targeted removal from the global model. The unlearning phase is executed on the server side, without requiring the target client to be online or maintaining any historical record of client participation. In the reported experimental results, we show that our method can achieve competitive or superior unlearning performance compared to state-of-the-art baselines, in particular in terms of more precise forgetting, using a very small distilled dataset (e.g., distilling only five data points per class).