Navigating Adversarial Robustness in Multimodal Systems

Multimodal systems that integrate vision, language, and other data modalities are becoming increasingly essential, raising emerging concerns about their adversarial robustness in real-world environment. This talk will delve into unique risks and complexities posed by adversarial threats in multimodal contexts, examining how these attacks disrupt model behaviour, and uncovering properties that inform more robust learning. We explore key challenges including intrinsic intra- and inter-modal interactions, overfitting and transferability, while gaining insights into strategies like multi-modal adversarial training, multi-modal interactions disruption, hierarchical utility enhancement, ensemble learning and data augmentation. These contribute to the development of systems that are not only accurate but also robust against a wide spectrum of adversarial threats.