Unleashing Uncertainty: Efficient Machine Unlearning for Generative AI

This paper is an extension of our CVPR2025 machine unlearning paper. There, we proposed an information-theoretical framework in an attempt to discriminate the information to be unlearned and the one to be retained, and a novel method, called LoTUS, that succeeded effective and efficient unlearning by increasing the model’s uncertainty about the data to be forgotten. Our previous work focused on classifiers.

In this paper we are presenting a brief overview of our novel method, SAFEMax that is guided by the information-theoretical foundation we have developed for machine unlearning. Preliminary experiments demonstrate that SAFEMax:

• succeeds significantly more efficient unlearning than existing methods,
• leverages the internal mechanisms and attributes of DDPMs in an attempt to minimize the priors needed for unlearning.

The research questions that arise are:

• Do we really need to anchor forgotten concepts/classes to other ones?
• Does unleashing uncertainty enhances the robustness of unlearning, in terms of resilience to unlearning attacks that attempt to restore the forgotten information?

Explore More: Poster

Recommended citation: Spartalis et al. "Unleashing Uncertainty: Efficient Machine Unlearning for Generative AI." ICML 2025 Workshop on Machine Unlearning for Generative AI. 2025.
Download Paper | Download Bibtex