Purpose
LLM-security-and-privacy is an open-source repository organized as a curated list, focusing on the security and privacy issues of large language models (LLMs). It is aimed at readers who want to understand LLM threats and vulnerabilities from both security and privacy perspectives, gathering related papers, tools, frameworks, and news for centralized reference and tracking.
Main Capabilities
The content organized in the repository includes:
- Paper list: covers research related to LLM threats and vulnerabilities, and is accompanied by an overview image showing the taxonomy of security and privacy threats against deep learning models and large language models.
- Paper summaries: in the paper-summaries folder, summaries, key points, and supplementary details are provided for each paper.
- Citation information: the main.bib file contains the latest citations for the papers listed in the list.
- Frameworks and taxonomies: lists reference frameworks and classification systems in the field of LLM security, such as OWASP Top 10 for Large Language Model Applications, MITRE ATLAS, NIST AI 100-2 E2023, etc.
- Tools and news: lists related tools (such as Arsenal - CALDERA Plugin, Vigil - Prompt injection detection) as well as news articles, blogs, and talks.
How to Use
Users can navigate to sections such as papers, frameworks and taxonomies, tools, and news articles through the Table of Contents on the GitHub repository page. To read a summary of an individual paper, users can enter the paper-summaries folder; to cite references, users can refer to main.bib. The repository also provides CONTRIBUTING.md with contribution guidelines and lists the current contributors.
Applicable Scenarios
Suitable for researchers, security engineers, and developers in the field of LLM security and privacy as a literature index and threat classification reference; it can also be used to learn about related tools and industry frameworks such as prompt injection detection.
Limitations
The repository itself is a collection of resource lists and summaries, and does not provide model training, deployment, or protection implementations; specific features, fees, prices, platform support, and other information are not currently provided.
