Research map / Adversarial robustness and security / Attacks on text and graphs
SafeLogo: Turning Your Logos into Jailbreak Shields via Micro-Regional Adversarial Training
Zhiyi Duan, Xiaoyue Zhang, Tianxing Man
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SafeLogo on top of a safety instruction (ablation) | LLaVA-1.5-13B · MM-SafetyBench SD+TYPO GCG | ASR (%): 12.8% -> 3.1% | Improved |
| use SafeLogo alone without safety instruction (ablation) | LLaVA-1.5-13B · MM-SafetyBench SD+TYPO GCG | ASR (%): 17.1% -> 18.7% | No clear effect |
| increase defense weight from 50% to 98% (ablation) | SafeLogo · MM-SafetyBench / MM-Vet | ASR (%) / MM-Vet: ASR 31% -> 6%; MM-Vet 39.9 -> 35.3 | Mixed |
| raise defense weight to 99% (ablation) | SafeLogo · MM-SafetyBench | ASR (%): 6% -> 8% | Hurt |
| increase logo coverage from 1% to 9% (ablation) | SafeLogo · MM-SafetyBench | ASR (%): 29% -> 5% | Improved |
| increase coverage from 9% to 16% (ablation) | SafeLogo · MM-SafetyBench / MM-Vet | ASR (%) / MM-Vet: ASR 5% -> 8%; MM-Vet 38.5 -> 37.0 | Hurt |
| perturbation range 64/255 versus 96/255 (ablation) | SafeLogo · MM-SafetyBench / MM-Vet | ASR (%) / MM-Vet: ASR 8% -> 15%; MM-Vet 39.1 -> 36.5 at 96/255 | Hurt |
| perturbation range 128/255 versus 64/255 (ablation) | SafeLogo · MM-SafetyBench / MM-Vet | ASR (%) / MM-Vet: ASR 8% -> 7%; MM-Vet 39.1 -> 37.2 | Mixed |
| full-intensity perturbation 255/255 (ablation) | SafeLogo · MM-SafetyBench | ASR (%): 8% -> 22% | Hurt |
| apply LLaVA-trained SafeLogo to Gemini-2.5-Flash-Image | SafeLogo · VLGuard (Gemini-2.5-Flash-Image) | ASR (%): 2.9/4.8/1.0% -> 1.0/0.0/0.0% | Improved |
| apply SafeLogo to Qwen3-VL-8B | Qwen3-VL-8B-Instruct · MM-Vet total | score: 65.8 -> 53.3 | Hurt |
| apply SafeLogo to LLaVA-1.5-13B | LLaVA-1.5-13B · MME total (OOD) | score: 1820 -> 1553 | Hurt |
Closest papers by meaning
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- Exploring Visual Vulnerabilities via Multi-Loss Adversarial Search for Jailbreaking Vision-Language Models (CVPR 2025 · similarity 0.90)
- Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks (CVPR 2025 · similarity 0.90)
- When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models (ICML 2026 · similarity 0.89)
- Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models (ICLR 2024 · similarity 0.89)
- Visual Adversarial Examples Jailbreak Aligned Large Language Models (AAAI 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
