Robustness of Simple AI Models in the Context of Machine Unlearning
Date:
Machine unlearning aims to give individuals data autonomy by removing their data’s influence from trained models. In this talk, I showed that these unlearning updates can paradoxically enable high-accuracy reconstruction attacks that recover the deleted data, especially for very simple models like linear regression. I also showed the generalization of these attacks across losses, architectures, and datasets, demonstrating that data deletion can introduce serious privacy risks even in settings previously considered safe.