Feature-indistinguishable machine unlearning via negative-hot label encoding and class weight masking

Abstract

Abstract With the growing importance of data privacy and regulatory compliance, machine unlearning has become a critical requirement in deep learning. However, existing approaches often require access to the original training data, incur substantial computational costs, or compromise performance on retained data. To address these limitations, we propose a novel unlearning framework that integrates label encoding fine-tuning with class weight masking, enabling efficient and selective forgetting of specific classes. In particular, we introduce Negative-Hot Label Encoding (NHLE), which suppresses the discriminability of target classes in the feature space, thereby weakening their representations. Our method requires only a small number of samples from the forgotten classes for iterative fine-tuning. Extensive experiments on multiple visual datasets show that the proposed framework achieves near-zero classification accuracy on forgotten data, while reducing accuracy on retained data by no more than 0.035.