Federated Knowledge Editing for LLMs
Privacy-preserving collaborative model updates without retraining
FLEKE introduces a federated approach for updating knowledge in large language models by enabling multiple organizations to collaboratively edit model knowledge while preserving privacy.
- Eliminates redundant computations when multiple clients update overlapping knowledge
- Preserves data privacy for sensitive information across decentralized organizations
- Achieves comparable editing performance to traditional methods without full model retraining
- Demonstrates 91.7% precision while maintaining overall model capabilities
This research is particularly valuable for medical institutions that need to collectively update medical knowledge in LLMs while maintaining patient data privacy and regulatory compliance.