Application of Secure Visual Language Model-Based Framework for Collaborative Analysis of Dermatological Images

Ensuring secure and privacy-preserving analysis of medical data remains a critical challenge in healthcare, particularly when sensitive patient information is distributed across institutions. This paper applies a secure visual-language framework designed for collaborative analysis of dermatological images, focusing on the DermaMNIST classification task. Leveraging vertical federated learning and the CLIP visual-language model, our framework enables distributed institutions to collaboratively train models without sharing raw data. By integrating prompt tuning, the framework adapts CLIP to medical image classification while maintaining data privacy. Experimental results demonstrate state-of-the-art performance on DermaMNIST, achieving an AUC of 0.933 and accuracy of 0.793, outperforming existing baselines.