MLLM

An adversarial attack that recovers supposedly unlearned multi-modality knowledge from MLLMs via prompt-suffix optimization and fine-tuning, exposing vulnerabilities in machine unlearning defenses.