ARC@ORU: Machine Unlearning: The Good, The Bad and The Ugly

ARC@ORU Research Seminar Series

Machine Unlearning: The Good, The Bad and The Ugly

About the talk

Machine unlearning is a process of removing the data subset, or the knowledge learned from it, from a trained machine learning (ML) model. This concept has become progressively more significant, as it supports the exercise of the right to be forgotten and facilitates compliance with privacy regulations such as the GDPR and the EU AI Act. Machine unlearning may seem elegant at first, but it comes with serious drawbacks. The proof of unlearning cannot be made solely by comparing output metrics or individual points in the parameter space. The model owner must provide verifiable evidence that the unlearning process has been carried out correctly. That usually means revealing model parameters and training history, yet doing so risks exposing sensitive model details and compromising model privacy. This tension creates a dangerous gap: Untrusted model owners can easily pretend they have unlearned data, and without a strong verification mechanism, catching this kind of deception is nearly impossible.

In this talk, Buse Atli will unpack both the promise and the pitfalls of machine unlearning, explain why current verification mechanisms are fundamentally broken, and explore what we can do to fix them so that ML systems can become genuinely more transparent and trustworthy.

Speaker

Buse Atli is an Assistant Professor in the Cybersecurity Division at Linköping University. Before joining LiU, she was a security researcher at Nokia Bell Labs, specializing in threat modeling strategies to address security and privacy challenges in integrating AI into network systems. Her current research focuses on trustworthy machine learning, including robustness, data privacy, model confidentiality, verifiability, and AI governance.