This research addresses the challenge of deploying large vision-language models (like CLIP) in distributed healthcare environments through a novel federated learning approach.

• Federated Adversarial Adaptation technique significantly reduces model size while maintaining performance
• Effectively handles data heterogeneity across different medical clients
• Demonstrates improved generalization performance on medical imaging datasets
• Preserves privacy by keeping sensitive medical data on local devices

This research enables healthcare organizations to leverage powerful vision-language models across distributed systems while addressing critical concerns around data privacy, computational efficiency, and performance in diverse medical contexts.

FAA-CLIP: Federated Adversarial Adaptation of CLIP