X-Domain Federated Meta: Transformer-Based Cross-Domain Federated Meta-Learning for Unified Medical Oncology and Agricultural Pathology Diagnostics
Conference paper presented at the International Mastermind Meet on Multidisciplinary Research and Innovation for Viksit Bharat, Pimpri Chinchwad University, Pune, Maharashtra, India.
- Introduced XFM, a privacy-preserving framework for cross-domain federated meta-learning across medical oncology and agricultural pathology.
- Used a shared ViT-Base/16 backbone with separate domain-specific classification heads to avoid destructive gradient interference.
- Enabled new clients to adapt from as few as 50 samples with a MAML-based meta-learning outer loop and applied differential privacy plus secure aggregation for compliance.
- Included Attention Rollout for diagnostic heatmaps at no extra inference cost, with validation planned at full scale.