多模態融合對抗攻擊

Risk Description

Attackers inject small but coordinated perturbations across multiple input modalities simultaneously (e.g., text+image, audio+visual), exploiting cross-modal fusion mechanisms to amplify attack effects. While single-modality perturbations may fall within safety thresholds, coordinated multi-modal perturbations can trigger severe model failures after fusion. As multi-modal AI systems (GPT-4V, Gemini) proliferate, this attack surface expands dramatically.