Provable Defense Against Adversarial Attacks on Multi-modal Models

We propose MMCert, the first certified defense against adversarial attacks on multi-modal models. MMCert provides provable robustness guarantees by leveraging a modality-independent sub-sampling strategy.