Continuous Technologies International Ltd
This project addresses critical security challenges posed by the proliferation of Large Language Models (LLMs) in various sectors. Our goal is to help organizations develop customizable protective measures to facilitate the secure usage of LLMs, particularly in mitigating sensitive data leakage and undesirable responses. The project will develop an LLM Access Shield system, which is an input-output security safeguard framework for sensitive data detection in user prompts and undesirable content detection in LLM responses, mainly composed of a Training Data Generation Module, a Domain-specific LLM for Security (DLMS) and a Sensitive Data Anonymization Module. The Training Data Generation Module is used to produce training data based on user-configurable sensitive categories and prompt templates. The DLMS is designed to acquire domain-specific vocabulary and context, enhancing its ability to identify potential data policy violations and undesired LLM responses via model fine-tuning. The Sensitive Data Anonymization Module facilitates the context to remain meaningful in the LLM responses through in-place Format Preserving Encryption (FPE) within user prompts. LLM Access Shield enables organizations to engage with LLMs securely, mitigating legal and financial risks, facilitating compliance with data protection laws, and enhancing overall trust in LLM-generated content.
