DP-DyLoRA introduces a breakthrough approach for fine-tuning large language models with strong privacy guarantees on resource-constrained devices.

• Combines differential privacy with dynamic low-rank adaptation to protect user data while maintaining model utility
• Achieves up to 13× parameter reduction compared to standard fine-tuning methods
• Demonstrates superior performance over existing differentially private federated learning approaches
• Enables practical deployment on mobile devices with limited computational resources

This innovation addresses critical security concerns in federated learning by preventing sensitive information leakage while making privacy-preserving AI accessible for real-world applications.

DP-DyLoRA: Fine-Tuning Transformer-Based Models On-Device under Differentially Private Federated Learning using Dynamic Low-Rank Adaptation