FirewaLLM: A Portable Data Protection and Recovery Framework for LLM Services

In the era characterized by the swift proliferation of large language models such as ChatGPT and GPT-4, there is a mounting escalation of apprehension regarding user privacy. These large language models possess the potential to inadvertently expose sensitive information, encompassing personal identities, health particulars, and financial data. The inadvertent exposure and misuse of such information can lead to significant privacy breaches, thereby exposing model owners to potential legal ramifications. This emphasizes the imperative necessity to amplify efforts in enhancing and evaluating data privacy and security protocols within the domain of large language models. Remarkably, a comprehensive framework for safeguarding user security and privacy is presently absent, leaving a discernible void in established standards for evaluating the privacy and security aspects of Big Predictive Models. To address this gap, we have proposed FirewaLLM, a portable framework that aims to protect user data security within the realm of Large Language Model services. This framework is specifically designed to encompass data protection and recovery measures, mitigating potential vulnerabilities and enhancing overall privacy safeguards. Within this framework, users employ a smaller model to locally desensitize sensitive aspects of text before submitting it to the large language model. By adopting this approach, privacy concerns are addressed proactively, as potentially identifying information is obfuscated prior to interacting with the large language model. Subsequently, the responses obtained from the large language model are matched with the original local text, facilitating the restoration of private information. This process ensures that the desired output is generated while preserving the confidentiality of sensitive data. Furthermore, we have introduced a bespoke benchmark specifically designed to evaluate the security and accuracy of large language models. This benchmark provides a comprehensive assessment of Large Language Models from two key perspectives: security and accuracy. Leveraging this benchmark, we have conducted a detailed evaluation and analysis of the security attributes of our local text desensitization tool in conjunction with ChatGPT-3.5. In conclusion, our research endeavors to tackle the pressing privacy concerns associated with large language models, providing a robust safeguard for user data and presenting a practical approach to evaluating the performance of these models, by employing a relatively smaller model for local desensitization. We believe that this study holds significant practical implications for upholding user privacy and data security within the context of LLM services. FirewaLLM is publicly released at https://github.com/ysy1216/FirewaLLM .