DomainShield

Privacy protection for LLM pipelines

DomainShield is a research project focused on preventing sensitive data leakage when using external large language model APIs.

Overview

The system acts as a middleware firewall:

• Masks sensitive information before sending data to external LLMs
• Handles both general PII and domain-specific sensitive entities
• Reconstructs the original content after receiving the response

Key Focus

• PII masking (names, emails, identifiers)
• Domain-specific entity protection (internal terms, codes, private vocabularies)
• Multilingual robustness under noisy conditions
• Comparison of adaptation methods (prompting, RAG, fine-tuning, NER)

Approach

We evaluate multiple strategies for detecting and masking sensitive data:

• Prompt-based methods
• Retrieval-augmented approaches (RAG)
• Supervised fine-tuning (LoRA)
• Token classification (NER)
• Hybrid and ensemble methods

Status

Active research project. Models, benchmarks, and demos coming soon.