Rethinking Backdoor Adversarial Unlearning through the Lens of Catastrophic Forgetting in Continual Learning

In three linesNovel approach to eliminate backdoors in AI models through continual learning lens. BI-BAU formulates unlearning as blind inversion solved via bi-level optimization and EM algorithm. Completely removes backdoor effects across targeted and untargeted attacks, including multi-modal contrastive learning.

Summary generated by Claude — human-verified