Adversarial Robustness Of Multimodal Machine Learning Models

The widespread adoption of machine learning, particularly generative models, has revolutionized productivity and capabilities. However, this progress comes with significant security risks, as adversarial attacks can subtly manipulate model predictions through imperceptible perturbations. Multimodal machine learning models, integrating data from modalities such as images and text, further expand the attack surface by enabling cross-modal exploitation. This paper examines the adversarial robustness of multimodal AI systems, focusing on text and image modalities. We analyze security challenges arising from cross-modal interactions and explore adversarial example generation techniques targeting individual modalities and shared embedding spaces. The paper concludes by identifying critical open challenges and promising research directions to enhance the security of multimodal AI systems.