LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty

We present a novel Information-Theoretic framework for Machine Unlearning—a principled method that enables AI models to forget specific training samples by carefully increasing uncertainty around them.

Unlike prior approaches that didn’t explore the information-theoretic nature of unlearning, our method uses this framework to discriminate between information that should be forgotten and information that should be retained, preserving the model’s overall utility while effectively removing targeted data.

What’s New:

• A novel Information-Theoretic framework for defining and guiding machine unlearning.
• A method that meticulously increases uncertainty only where needed—efficient and utility-preserving.
• A new real-world-focused metric for evaluating unlearning performance in practical deployment scenarios.

Explore More: Poster | Code & Overview

Recommended citation: Spartalis et al. "LoTUS: Large-Scale Machine Unlearning with a Taste of Uncertainty." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2025.
Download Paper | Download Slides | Download Bibtex