Weight masking in image classification networks: class-specific machine unlearning

Jiali Wang, Hongxia Bie, Zhao Jing, Yichen Zhi, Yongkai Fan

Abstract: Machine unlearning focuses on data forgetting, rendering the model unable to recognize previously identifiable data, which holds promise for ensuring data privacy and security. However, current methods are computationally intensive and can negatively impact classification performance on the retained data. In this paper, we propose a new unlearning method that can blur specific classes via weight masking, without retraining and additional computation. The weights, specific to each class by mapping the feature vectors of corresponding class samples, represent the statistics of these feature vectors and indicate the importance distribution of feature activation units for each class. This method has been validated on multiple datasets, showing a negligible ability to recognize forgotten data, with accuracy dropping to 0–0.32%, while increasing by 0.08–18.85% compared to the original networks on retained data.

External IDs:dblp:journals/kais/WangBJZF25