Rethinking Federated Unlearning via the Lens of Memorization
Researchers have proposed a new method for federated unlearning, a process crucial for complying with privacy regulations in machine learning. Their approach, called Federated Memorization Pruning (FedMemPrune), focuses on removing uniquely memorized information from specific data points rather than general knowledge shared across datasets. This method uses a novel metric, Grouped Memorization Evaluation, to distinguish between memorized and overlapping information. Experiments indicate that FedMemPrune effectively eliminates memorization while preserving the utility of the remaining data, matching the performance of retraining-based methods. AI
IMPACT Introduces a novel approach to data privacy in federated learning, potentially improving compliance and model utility.