thuccslab / awesome-lm-ssp

A reading list for large models safety, security, and privacy.

Awesome-LM-SSP

Introduction

The resources related to the trustworthiness of large models (LMs) across multiple dimensions (e.g., safety, security, and privacy), with a special focus on multi-modal LMs (e.g., vision-language models and diffusion models).

• This repo is in progress 🌱 (currently manually collected).

• Badges:

  • Model:

  • Comment:

  • Venue (Continuous update): or

• 🌻 Welcome to recommend resources to us via Issues with the following format (please fill in this table):

Title	Link	Code	Venue	Classification	Model	Comment
aa	arxiv	github	bb'23	A1. Jailbreak	LLM	Agent

News

• [2024.04.27] We adjusted the categories.
• [2024.01.20] We collected 3 related papers from NDSS'24!
• [2024.01.17] We collected 108 related papers from ICLR'24!
• [2024.01.09] 🚀 LM-SSP is released!

Collections

• Book (1)
• Competition (5)
• Leaderboard (3)
• Toolkit (6)
• Survey (23)
• Paper (710)
  • A. Safety (435)
    • A0. General (6)
    • A1. Jailbreak (134)
    • A2. Alignment (47)
    • A3. Deepfake (36)
    • A4. Ethics (5)
    • A5. Fairness (49)
    • A6. Hallucination (97)
    • A7. Prompt Injection (13)
    • A8. Toxicity (48)
  • B. Security (115)
    • B0. General (1)
    • B1. Adversarial Examples (61)
    • B2. Poison & Backdoor (46)
    • B3. System (7)
  • C. Privacy (160)
    • C0. General (14)
    • C1. Contamination (8)
    • C2. Copyright (52)
    • C3. Data Reconstruction (18)
    • C4. Membership Inference Attacks (9)
    • C5. Model Extraction (7)
    • C6. Privacy-Preserving Computation (22)
    • C7. Unlearning (30)

Star History

Acknowledgement

• Organizers: Tianshuo Cong (丛天硕), Xinlei He (何新磊), Zhengyu Zhao (赵正宇), Yugeng Liu (刘禹更), Delong Ran (冉德龙)

• This project is inspired by LLM Security, Awesome LLM Security, LLM Security & Privacy, UR2-LLMs, PLMpapers, EvaluationPapers4ChatGPT

awesome-lm-ssp's Issues

What is the difference between Data Reconstruction and Extraction?

我认为Data Reconstruction是指从公共聚合信息中，部分重建私有数据集的方法。比如基于开源语言模型，加入私有数据进行训练。对私有数据的攻击是Data Reconstruction（刚接触这个领域，不知道这样描述对不对）。可是在Data Reconstruction中看到了[Extracting Training Data from Large Language Models]这篇文章。

request for adding a new survey

Hi, thank you for this great repo. Could you please add this new survey? Attacks, Defenses and Evaluations for LLM Conversation Safety: A Survey

Kindly request the inclusion

Title	Link	Code	Venue	Classification	Model	Comment
AdaShield: Safeguarding Multimodal Large Language Models from Structure-based Attack via Adaptive Shield Prompting	https://arxiv.org/abs/2403.09513	https://github.com/rain305f/AdaShield	arXiv'24	A1. Jailbreak	VLM	VLM Jailbreak Defense

Some of my related works

Title	Link	Code	Venue	Classification	Model	Comment
Towards More Effective Protection Against Diffusion-Based Mimicry with Score Distillation	https://arxiv.org/abs/2311.12832	https://github.com/xavihart/Diff-Protect	ICLR 2024	C2. Copyright	Diffusion Model	protective perturbation of diffusion model
Diffusion-Based Adversarial Sample Generation for Improved Stealthiness and Controllability	https://arxiv.org/abs/2305.16494	https://github.com/xavihart/Diff-PGD	NeurIPS 2023	B1. Adversarial Samples	Diffusion Model	generate stealthy adversarial samples

Kindly request the inclusion

Title	Link	Code	Venue	Classification	Model	Comment
Automatic and Universal Prompt Injection Attacks against Large Language Models	https://arxiv.org/abs/2403.04957	https://github.com/SheltonLiu-N/Universal-Prompt-Injection	arXiv'24	A7. Prompt Injection	LLM	Automatically generating highly effective and universal prompt injection data

Some works that have not been included

👍 Thank you for creating and maintaining such a great repository. I found that these works have not been included and hope they can be added.

Title	Link	Code	Venue	Classification	Model	Comment
Query-Relevant Images Jailbreak Large Multi-Modal Models	https://arxiv.org/abs/2311.17600	https://github.com/isXinLiu/MM-SafetyBench	arXiv'23	A1. Jailbreak	VLM
GUARD: Role-playing to Generate Natural-language Jailbreakings to Test Guideline Adherence of Large Language Models	https://arxiv.org/abs/2402.03299		arXiv'24	A1. Jailbreak	LLM
On the Robustness of Large Multimodal Models Against Image Adversarial Attacks	https://arxiv.org/abs/2312.03777		arXiv'23	B1. Adversarial Examples	VLM
VL-Trojan: Multimodal Instruction Backdoor Attacks against Autoregressive Visual Language Models	https://arxiv.org/abs/2402.13851		arXiv'24	B2. Poisoning	VLM