Assessing the Adversarial Robustness of Multimodal Medical AI Systems: Insights into Vulnerabilities and Modality Interactions

Ekaterina Mozhegova^1*Asad Masood Khattak²Adil Khan³Roman Garaev¹Bader Rasheed¹

• ¹Innopolis University, Innopolis, Russia
• ²Zayed University, Abu Dhabi, United Arab Emirates
• ³University of Hull, Hull, Yorkshire and the Humber, United Kingdom

Deep learning systems have been expanding their potential to be used with different data types a4 a wide range of fields. Medicine is one of the most compelling examples of AI applications due to the numerous tasks and various data types associated with them, such as texts, images, and numerical records. Both task-specific single-modality models and generalpurpose multimodal large models have emerged, presenting new opportunities along with new challenges. Despite the potential of AI systems, regardless of the task, they remain vulnerable to adversarial attacks. These attacks are small perturbations, often imperceptible to the human eye, that are intentionally generated to deceive the model. In high-stakes domains, including healthcare as one of the notable examples, they pose significant risks to model's reliability and consequently to their prospects of being applied in real-world tasks. This is why studies focused on adversarial robustness are of such great importance.In our research, we contribute to the fields of multimodality and adversarial robustness, with a specific focus on applications in medicine. As we investigate the behavior of multimodal models under various attack scenarios, this research sheds light on how these models endure attacks and whether the multimodal nature of these models enhances their resilience. We conducted experiments by applying attacks to two modalities: images and texts. We applied attacks to single-modality models, combined these models with modality fusion techniques and applied the same attacks on the combined model. This approach allowed us to validate our hypothesis that multimodality enhances robustness to adversarial attacks. We believe that our findings have the potential to open new horizons in the field of robust multimodality systems. Additionally, this study can initiate further research focused on the data flow in multimodal systems.