Research map / Adversarial robustness and security / Attacks on text and graphs
SafetyReminder: Reviving Delayed Safety Awareness of Vision-Language Models to Defend Against Jailbreak Attacks
Peiyuan Tang, Haojie Xin, Xiaodong Zhang, Jun Sun, Qin Xia, Zijiang James Yang
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 classification loss Lcls to generation loss (ablation) | SAPT soft prompt · LLaVA-1.5-7B avg ASR | attack success rate: 19.7 -> 6.2 | Improved |
| add generation loss Lgen to Lcls (ablation) | SAPT soft prompt · LLaVA-1.5-7B avg ASR | attack success rate: 74.7 -> 6.2 | Improved |
| apply SAPT defense | LLaVA-1.5-7B · avg ASR over 6 benchmarks/attacks | attack success rate: 76.1 -> 5.0 | Improved |
| apply SAPT defense | Qwen2-VL-7B · avg ASR | attack success rate: 50.0 -> 4.7 | Improved |
| apply SAPT defense | LLaVA-1.5-7B · MM-Vet utility | score: 27.9 -> 28.0 | No clear effect |
| apply SAPT defense | LLaVA-1.5-7B · MM-Vet refusal rate | refusal rate: 2.3 -> 3.7 | Hurt |
| apply SAPT instead of ASTRA | Qwen2-VL-7B · ImgJP | attack success rate: 2.8 -> 8.0 | Hurt |
| use Self-Reminder/DRO prompt defenses against ImgJP | Qwen2-VL-7B · ImgJP | attack success rate: 97.4 -> 95.4 | No clear effect |
| lengthen soft prompt beyond 4 tokens (ablation) | SAPT · LLaVA-1.5-7B MM-Vet | score | Hurt |
| increase safety check interval (ablation) | SAPT · LLaVA-1.5-7B | attack success rate | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- BlueSuffix — BlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.92)
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility (ICML 2026 · similarity 0.92)
- Jailbreaking Vision-Language Models Through the Visual Modality (ICML 2026 · similarity 0.92)
- Safe + Safe = Unsafe? Exploring How Safe Images Can Be Exploited to Jailbreak Large Vision-Language Models (NeurIPS 2025 · similarity 0.91)
- Retention Score: Quantifying Jailbreak Risks for Vision Language Models (AAAI 2025 · similarity 0.91)
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore Mechanism (NeurIPS 2025 · similarity 0.91)
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.
