# PromptSmooth: Certifying Robustness of Medical Vision-Language Models via Prompt Learning

**Authors:** Noor Hussein, Fahad Shamshad, Muzammal Naseer, Karthik Nandakumar  
**arXiv:** [2408.16769](https://arxiv.org/abs/2408.16769)  
**Date:** 2024-08-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.16769v1/

## Key points

- Medical Vision-Language Models (Med-VLMs) are widely used in processing medical image-text pairs
- Med-VLMs are vulnerable to adversarial attacks, raising concerns about their reliability and robustness
- PromptSmooth is a novel framework designed to enhance the certified robustness of Med-VLMs by leveraging prompt learning techniques
- PromptSmooth adapts pre-trained Med-VLMs to handle Gaussian noise through the learning of textual prompts in a zero-shot or few-shot manner
- PromptSmooth achieves a balance between accuracy and robustness while minimizing computational overhead
- It only requires a single model to handle multiple noise levels, reducing computational costs compared to traditional methods
- PromptSmooth outperformed existing approaches in terms of both performance and practicality through comprehensive experiments involving three different Med-VLMs and six downstream datasets representing various imaging modalities
- It does not require extensive medical datasets, enhancing its applicability in real-world scenarios
- Overall, PromptSmooth represents a significant advancement in ensuring the robustness of Med-VLMs against adversarial attacks by offering an innovative approach to prompt learning.

## Summary

In the field of medical image analysis, Medical Vision-Language Models (Med-VLMs) have become a popular tool for processing large datasets of medical image-text pairs. Recent studies have shown that these models are vulnerable to adversarial attacks, raising concerns about their reliability and robustness. To address this issue, researchers have proposed a novel framework called PromptSmooth. <br>
PromptSmooth is designed to enhance the certified robustness of Med-VLMs by leveraging prompt learning techniques. By adapting pre-trained Med-VLMs to handle Gaussian noise through the learning of textual prompts in a zero-shot or few-shot manner, PromptSmooth achieves a delicate balance between accuracy and robustness while minimizing computational overhead. Importantly, PromptSmooth only requires a single model to handle multiple noise levels, reducing computational costs compared to traditional methods that rely on training separate models for each noise level.<br>
The efficiency and effectiveness of PromptSmooth were demonstrated through comprehensive experiments involving three different Med-VLMs and six downstream datasets representing various imaging modalities. The results showed that PromptSmooth outperformed existing approaches in terms of both performance and practicality. Additionally, PromptSmooth does not require extensive medical datasets, further enhancing its applicability in real-world scenarios.<br>
Overall, PromptSmooth represents a significant advancement in ensuring the robustness of Med-VLMs against adversarial attacks. Its innovative approach to prompt learning offers a promising solution for enhancing the security and reliability of medical image analysis systems.