Jaewon Lee, Yongwoo Kim, Donghyun Kim
The Erase at the Core (EC) framework enhances machine unlearning by ensuring that forgetting occurs throughout the entire network, not just at the final classifier level, leading to better representational forgetting while preserving performance on retained data.
Machine unlearning is about making a machine learning model forget certain data, but many current methods only make it appear as if data is forgotten without actually erasing the underlying information. This paper introduces a new approach called Erase at the Core (EC) that ensures data is truly forgotten throughout the model, not just superficially. EC does this by applying special techniques to all layers of the model, making it better at forgetting specific data while still working well with the data it needs to remember. This approach can be added to existing unlearning methods to improve their effectiveness.
