diff --git a/README.md b/README.md index d9c5fc2..16b5bec 100644 --- a/README.md +++ b/README.md @@ -1,40 +1,72 @@ -# OmniSafeBench-MM -A Unified Benchmark and Toolbox for Multimodal Jailbreak Attack–Defense Evaluation +

+ +OmniSafeBench-MM + +

+ +
+ +# 🗡️ OmniSafeBench-MM: A Unified Benchmark and Toolbox for Multimodal Jailbreak Attack–Defense Evaluation 🛡️ + +
+ +## 📰 News 📰 +* ***[2024/02/06] 🚀 [Initial release of OmniSafeBench-MM with 13 attack methods and 15 defense strategies]()*** 🚀 + + + + +## 🗡️ What is OmniSafeBench-MM 🛡️ +OmniSafeBench-MM is a unified benchmark and open-source toolbox for evaluating multimodal jailbreak attacks and defenses in Large Vision–Language Models (MLLMs). It integrates a large-scale dataset spanning 8–9 major risk domains and 50 fine-grained categories, supports three real-world prompt types (consultative, imperative, declarative), and implements 13 representative attack methods and 15 defense strategies in a modular pipeline. Beyond traditional ASR, it introduces a three-dimensional evaluation protocol measuring harmfulness, intent alignment, and response detail, enabling fine-grained safety–utility analysis. Tested across 18 open-source and closed-source MLLMs, OmniSafeBench-MM provides a comprehensive, reproducible, and extensible platform for benchmarking multimodal safety. + + + +## 🌐 Overview 🌐 +
+ +HarmBench Evaluation Pipeline + +
+**Overview of OmniSafeBench-MM**. +The benchmark unifies multi-modal jailbreak attack–defense evaluation, 13 attack and 15 defense methods, and a three-dimensional scoring protocol measuring harmfulness, alignment, and detail. ## 🗡️ Integrated Attack Methods -| Name | Full paper title | Paper | Code | Classification | -|------|------------------|-------|------|----------------| -| FigStep / FigStep-Pro | FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts | [paper link](https://arxiv.org/abs/2311.05608) | [code link](https://github.com/ThuCCSLab/FigStep) | Black-box — Structured visual carriers | -| QR-Attack (MM-SafetyBench) | MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Models | [paper link](https://arxiv.org/abs/2311.17600) | [code link](https://github.com/isXinLiu/MM-SafetyBench) | Black-box — Structured visual carriers | -| MML | Jailbreak Large Vision-Language Models Through Multi-Modal Linkage | [paper link](https://aclanthology.org/2025.acl-long.74/) | [code link](https://github.com/wangyu-ovo/MML) | Black-box — Structured visual carriers | -| CS-DJ | Distraction is All You Need for Multimodal Large Language Model Jailbreaking | [paper link](https://arxiv.org/abs/2502.10794) | [code link](https://github.com/TeamPigeonLab/CS-DJ) | Black-box — OOD (attention / distribution manipulation) | -| SI-Attack | Jailbreaking Multimodal Large Language Models via Shuffle Inconsistency | [paper link](https://arxiv.org/abs/2501.04931) | [code link](https://github.com/zhaoshiji123/SI-Attack) | Black-box — OOD (shuffle / attention inconsistency) | -| JOOD | Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy | [paper link](https://arxiv.org/abs/2503.20823) | [code link](https://github.com/naver-ai/JOOD) | Black-box — OOD (OOD strategy) | -| HIMRD | Heuristic-Induced Multimodal Risk Distribution (HIMRD) Jailbreak Attack | [paper link](https://arxiv.org/abs/2412.05934) | [code link](https://github.com/MaTengSYSU/HIMRD-jailbreak) | Black-box — OOD / risk distribution | -| HADES | Images are Achilles’ Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models | [paper link](https://arxiv.org/abs/2403.09792) | [code link](https://github.com/AoiDragon/HADES) | Black-box — Structured visual carriers | -| BAP | Jailbreak Vision Language Models via Bi-Modal Adversarial Prompt (BAP) | [paper link](https://arxiv.org/abs/2406.04031) | [code link](https://github.com/NY1024/BAP-Jailbreak-Vision-Language-Models-via-Bi-Modal-Adversarial-Prompt) | White-box — Cross-modal | -| visual_adv | Visual Adversarial Examples Jailbreak Aligned Large Language Models | [paper link](https://ojs.aaai.org/index.php/AAAI/article/view/30150) | [code link](https://github.com/Unispac/Visual-Adversarial-Examples-Jailbreak-Large-Language-Models) | White-box — Single-modal | -| VisCRA | VisCRA: A Visual Chain Reasoning Attack for Jailbreaking Multimodal Large Language Models | [paper link](https://arxiv.org/abs/2505.19684) | [code link](https://github.com/DyMessi/VisCRA) | Black-box — Structured visual carriers | -| UMK | White-box Multimodal Jailbreaks Against Large Vision-Language Models (Universal Master Key, UMK) | [paper link](https://arxiv.org/abs/2405.17894) | [code link](https://github.com/roywang021/UMK) | White-box — Cross-modal | -| PBI-Attack | Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for Toxicity Maximization (PBI-Attack) | [paper link](https://aclanthology.org/2025.emnlp-main.32.pdf) | [code link](https://github.com/Rosy0912/PBI-Attack) | Black-box — Query optimization & transfer | -| ImgJP / DeltaJP | Jailbreaking Attack against Multimodal Large Language Model (imgJP / deltaJP) | [paper link](https://arxiv.org/abs/2402.02309) | [code link](https://github.com/abc03570128/Jailbreaking-Attack-against-Multimodal-Large-Language-Model) | White-box — Single-modal | -| JPS | JPS: Jailbreak Multimodal Large Language Models with Collaborative Visual Perturbation and Textual Steering | [paper link](https://arxiv.org/abs/2508.05087) | [code link](https://github.com/thu-coai/JPS) | White-box — Cross-modal | +| Name | Title | Paper | Code | Classification | +|------|-----------------------------------------------------------------------------------------------------------------------------|-------|------|----------------| +| FigStep / FigStep-Pro | FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts | [link](https://arxiv.org/abs/2311.05608) | [link](https://github.com/ThuCCSLab/FigStep) | Black-box — Structured visual carriers | +| QR-Attack (MM-SafetyBench) | MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Models | [link](https://arxiv.org/abs/2311.17600) | [link](https://github.com/isXinLiu/MM-SafetyBench) | Black-box — Structured visual carriers | +| MML | Jailbreak Large Vision-Language Models Through Multi-Modal Linkage | [link](https://aclanthology.org/2025.acl-long.74/) | [link](https://github.com/wangyu-ovo/MML) | Black-box — Structured visual carriers | +| CS-DJ | Distraction is All You Need for Multimodal Large Language Model Jailbreaking | [link](https://arxiv.org/abs/2502.10794) | [link](https://github.com/TeamPigeonLab/CS-DJ) | Black-box — OOD (attention / distribution manipulation) | +| SI-Attack | Jailbreaking Multimodal Large Language Models via Shuffle Inconsistency | [link](https://arxiv.org/abs/2501.04931) | [link](https://github.com/zhaoshiji123/SI-Attack) | Black-box — OOD (shuffle / attention inconsistency) | +| JOOD | Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy | [link](https://arxiv.org/abs/2503.20823) | [link](https://github.com/naver-ai/JOOD) | Black-box — OOD (OOD strategy) | +| HIMRD | Heuristic-Induced Multimodal Risk Distribution (HIMRD) Jailbreak Attack | [link](https://arxiv.org/abs/2412.05934) | [link](https://github.com/MaTengSYSU/HIMRD-jailbreak) | Black-box — OOD / risk distribution | +| HADES | Images are Achilles’ Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models | [link](https://arxiv.org/abs/2403.09792) | [link](https://github.com/AoiDragon/HADES) | Black-box — Structured visual carriers | +| BAP | Jailbreak Vision Language Models via Bi-Modal Adversarial Prompt (BAP) | [link](https://arxiv.org/abs/2406.04031) | [link](https://github.com/NY1024/BAP-Jailbreak-Vision-Language-Models-via-Bi-Modal-Adversarial-Prompt) | White-box — Cross-modal | +| visual_adv | Visual Adversarial Examples Jailbreak Aligned Large Language Models | [link](https://ojs.aaai.org/index.php/AAAI/article/view/30150) | [link](https://github.com/Unispac/Visual-Adversarial-Examples-Jailbreak-Large-Language-Models) | White-box — Single-modal | +| VisCRA | VisCRA: A Visual Chain Reasoning Attack for Jailbreaking Multimodal Large Language Models | [link](https://arxiv.org/abs/2505.19684) | [link](https://github.com/DyMessi/VisCRA) | Black-box — Structured visual carriers | +| UMK | White-box Multimodal Jailbreaks Against Large Vision-Language Models (Universal Master Key, UMK) | [link](https://arxiv.org/abs/2405.17894) | [link](https://github.com/roywang021/UMK) | White-box — Cross-modal | +| PBI-Attack | Prior-Guided Bimodal Interactive Black-Box Jailbreak Attack for Toxicity Maximization (PBI-Attack) | [link](https://aclanthology.org/2025.emnlp-main.32.pdf) | [link](https://github.com/Rosy0912/PBI-Attack) | Black-box — Query optimization & transfer | +| ImgJP / DeltaJP | Jailbreaking Attack against Multimodal Large Language Model (imgJP / deltaJP) | [link](https://arxiv.org/abs/2402.02309) | [link](https://github.com/abc03570128/Jailbreaking-Attack-against-Multimodal-Large-Language-Model) | White-box — Single-modal | +| JPS | JPS: Jailbreak Multimodal Large Language Models with Collaborative Visual Perturbation and Textual Steering | [link](https://arxiv.org/abs/2508.05087) | [link](https://github.com/thu-coai/JPS) | White-box — Cross-modal | ## 🛡️Integrated Defense Methods -| No. | Title | Type | Venue | Paper | Code | -|:---|:--------------------------------------------------------------------------------------------------------------------------------------------------------------------|:----:|:---:|:-----------|:-----------| -| 1 | Jailguard: A universal detection framework for llm prompt-based attacks | 1 | arXiv | [paper link](https://arxiv.org/pdf/2312.10766) | [code link](https://github.com/shiningrain/JailGuard) | -| 2 | Mllm-protector: Ensuring mllm's safety without hurting performance | 1 | EMNLP 2024 | [paper link](https://arxiv.org/pdf/2401.02906) | [code link](https://github.com/pipilurj/MLLM-protector) | -| 3 | Eyes closed, safety on: Protecting multimodal llms via image-to-text transformation | 1 | ECCV 2024 | [paper link](https://arxiv.org/pdf/2403.09572) | [code link](https://gyhdog99.github.io/projects/ecso/) | -| 4 | Shieldlm: Empowering llms as aligned, customizable and explainable safety detectors | 1 | EMNLP 2024 | [paper link](https://arxiv.org/pdf/2402.16444) | [code link](https://github.com/thu-coai/ShieldLM) | -| 5 | Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting | 1 | ECCV 2024 | [paper link](https://link.springer.com/content/pdf/10.1007/978-3-031-72661-3_5.pdf) | [code link](https://github.com/SaFoLab-WISC/AdaShield)| -| 6 | Uniguard: Towards universal safety guardrails for jailbreak attacks on multimodal large language models | 1 | ECCV 2024 | [paper link](https://arxiv.org/pdf/2411.01703) | [code link](https://anonymous.4open.science/r/UniGuard/README.md) | -| 7 | Defending LVLMs Against Vision Attacks Through Partial-Perception Supervision | 1 | ICML 2025 | [paper link](https://arxiv.org/pdf/2412.12722) | [code link](https://github.com/tools-only/DPS) | -| 8 | Cross-modality Information Check for Detecting Jailbreaking in Multimodal Large Language Models | 1 | EMNLP 2024 | [paper link](https://arxiv.org/pdf/2407.21659) | [code link](https://github.com/PandragonXIII/CIDER) | -| 9 | Guardreasoner-vl: Safeguarding vlms via reinforced reasoning | 1 | ICML 2025 | [paper link](https://arxiv.org/pdf/2505.11049) | [code link](https://github.com/yueliu1999/GuardReasoner-VL/) | -| 10 | Llama Guard 4 | 1 | None | [document link](https://www.llama.com/docs/model-cards-and-prompt-formats/llama-guard-4/) | [model link](https://huggingface.co/meta-llama/Llama-Guard-4-12B) | -| 11 | QGuard: Question-based Zero-shot Guard for Multi-modal LLM Safety | 1 | arXiv | [paper link](https://arxiv.org/pdf/2506.12299) | [code link](https://github.com/taegyeong-lee/QGuard-Question-based-Zero-shot-Guard-for-Multi-modal-LLM-Safety) | -| 12 | Llavaguard: Vlm-based safeguard for vision dataset curation and safety assessment | 1 | CVPR 2024 | [paper link](https://openaccess.thecvf.com/content/CVPR2024W/ReGenAI/papers/Helff_LLAVAGUARD_VLM-based_Safeguard_for_Vision_Dataset_Curation_and_Safety_Assessment_CVPRW_2024_paper.pdf) | [code link](https://github.com/ml-research/LlavaGuard) | - +| Name | Title | Venue | Paper | Code | Classification | +|---|--------------------------------------------------------------------------------------------------------------------|---|---|---|---| +| JailGuard | JailGuard: A Universal Detection Framework for Prompt-based Attacks on LLM Systems | arXiv | [link](https://arxiv.org/abs/2312.10766) | [link](https://github.com/shiningrain/JailGuard) | Off-Model — Input pre-processing | +| MLLM-Protector | MLLM-Protector: Ensuring MLLM's Safety without Hurting Performance | EMNLP 2024 | [link](https://arxiv.org/abs/2401.02906) | [link](https://github.com/pipilurj/MLLM-protector) | Off-Model — Output post-processing | +| ECSO | Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation | ECCV 2024 | [link](https://arxiv.org/abs/2403.09572) | [link](https://gyhdog99.github.io/projects/ecso/) | Off-Model — Input pre-processing | +| ShieldLM | ShieldLM: Empowering LLMs as Aligned, Customizable and Explainable Safety Detectors | EMNLP 2024 | [link](https://arxiv.org/abs/2402.16444) | [link](https://github.com/thu-coai/ShieldLM) | Off-Model — Output post-processing | +| AdaShield | AdaShield: Safeguarding Multimodal Large Language Models from Structure-based Attack via Adaptive Shield Prompting | ECCV 2024 | [link](https://arxiv.org/abs/2403.09513) | [link](https://github.com/rain305f/AdaShield) | Off-Model — Input pre-processing | +| Uniguard | UNIGUARD: Towards Universal Safety Guardrails for Jailbreak Attacks on Multimodal Large Language Models | ECCV 2024 | [link](https://arxiv.org/abs/2411.01703) | [link](https://anonymous.4open.science/r/UniGuard/README.md) | Off-Model — Input pre-processing | +| DPS | Defending LVLMs Against Vision Attacks Through Partial-Perception Supervision | ICML 2025 | [link](https://arxiv.org/abs/2412.12722) | [link](https://github.com/tools-only/DPS) | Off-Model — Input pre-processing | +| CIDER | Cross-modality Information Check for Detecting Jailbreaking in Multimodal Large Language Models | EMNLP 2024 | [link](https://arxiv.org/abs/2407.21659) | [link](https://github.com/PandragonXIII/CIDER) | Off-Model — Input pre-processing | +| GuardReasoner-VL | GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning | ICML 2025 | [link](https://arxiv.org/abs/2505.11049) | [link](https://github.com/yueliu1999/GuardReasoner-VL) | Off-Model — Input pre-processing | +| Llama-Guard-4 | Llama Guard 4 | Model Card | [link](https://www.llama.com/docs/model-cards-and-prompt-formats/llama-guard-4/) | [link](https://huggingface.co/meta-llama/Llama-Guard-4-12B) | Off-Model — Input pre-processing | +| QGuard | QGuard: Question-based Zero-shot Guard for Multi-modal LLM Safety | arXiv | [link](https://arxiv.org/abs/2506.12299) | [link](https://github.com/taegyeong-lee/QGuard-Question-based-Zero-shot-Guard-for-Multi-modal-LLM-Safety) | Off-Model — Input pre-processing | +| LlavaGuard | LlavaGuard: An Open VLM-based Framework for Safeguarding Vision Datasets and Models | ICML 2025 | [link](https://arxiv.org/abs/2406.05113) | [link](https://github.com/ml-research/LlavaGuard) | Off-Model — Input pre-processing | +| Llama-Guard-3 | Llama Guard 3 | Model Card | [link](https://www.llama.com/docs/model-cards-and-prompt-formats/llama-guard-3/) | [model link](https://huggingface.co/meta-llama/Llama-Guard-3-8B) | Off-Model — Output post-processing | +| HiddenDetect | HiddenDetect: Detecting Jailbreak Attacks against Multimodal Large Language Models via Monitoring Hidden States | ACL 2025 | [link](https://arxiv.org/abs/2502.14744) | [link](https://github.com/leigest519/HiddenDetect) | On-Model — Inference process intervention | +| CoCA | CoCA: Regaining Safety-awareness of Multimodal Large Language Models with Constitutional Calibration | COLM 2024 | [link](https://arxiv.org/abs/2409.11365) | Null | On-Model — Inference process intervention | +| VLGuard | Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models | ICML 2024 | [link](https://arxiv.org/abs/2402.02207) | [link](https://github.com/ys-zong/VLGuard) | On-Model — Intrinsic model alignment | > More methods are coming soon!! diff --git a/assets/home.png b/assets/home.png new file mode 100644 index 0000000..83480bf Binary files /dev/null and b/assets/home.png differ diff --git a/assets/overview.png b/assets/overview.png new file mode 100644 index 0000000..bf026d4 Binary files /dev/null and b/assets/overview.png differ