AdaShield: Safeguarding Multimodal Large Language Models from Structure-based Attacks

AdaShield defends MLLMs against structure-based jailbreak attacks without fine-tuning or additional training.