On the Robustness of Machine Unlearning for Vision-Language Models

Researchers released the first systematic survey and robustness study of machine unlearning for vision-language models.

They introduced three attack tests and found many unlearning methods can be bypassed by contextual prompting and retraining.

The results suggest these systems often hide memorized knowledge rather than fully erase it.

This raises concerns for privacy, safety, and governance in multimodal AI.