Research map / Adversarial robustness and security / Attacks on text and graphs
Principled Steering via Null-space Projection for Jailbreak Defense in Vision-Language Models
Xingyu Zhu, Beier Zhu, Shuo Wang, Junfeng Fang, Kesen Zhao, Hanwang Zhang, Xiangnan He
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 refusal-alignment term to the steering objective (ablation) | NullSteer (MiniGPT-4) · Jailbreak (perturbation, unconstrained) | ASR: 8.57 -> 7.32 | Improved |
| add harmful-suppression term to the steering objective (ablation) | NullSteer (MiniGPT-4) · Jailbreak (perturbation, unconstrained) | ASR: 8.36 -> 7.32 | Improved |
| add harmful-suppression term to the steering objective (ablation) | NullSteer (MiniGPT-4) · Toxicity (perturbation, unconstrained) | Toxicity Score: 3.58 -> 2.89 | Improved |
| add refusal-alignment term to the steering objective (ablation) | NullSteer (MiniGPT-4) · Toxicity (perturbation, unconstrained) | Toxicity Score: 4.02 -> 2.89 | Improved |
| add all three objective terms (vs none) (ablation) | NullSteer (MiniGPT-4) · Toxicity (perturbation, unconstrained) | Toxicity Score: 30.65 -> 2.89 | Improved |
| add null-space-constrained steering vs ASTRA | MiniGPT-4 · Jailbreak (perturbation, unconstrained) | ASR: 9.09 -> 7.32 | Improved |
| add null-space-constrained steering vs ASTRA | LLaVA-v1.5 · Toxicity (perturbation, eps=32/255) | Toxicity Score: 34.76 -> 31.82 | Improved |
| increase number of benign activations N_b (ablation) | NullSteer · MM-Vet and MMBench utility | Utility score | Improved |
| increase number of malicious activations N_m (ablation) | NullSteer (MiniGPT-4) · Toxicity and Jailbreak | Toxicity Score and ASR | Improved |
| increase steering strength lambda (ablation) | NullSteer (MiniGPT-4) · Toxicity and Jailbreak (eps=16/255) | Toxicity Score and ASR | Improved |
| increase steering strength lambda (ablation) | NullSteer (MiniGPT-4) · MM-Vet | Utility score | No clear effect |
| apply NullSteer to MiniGPT-4 vs vanilla | MiniGPT-4 · MM-Vet | Utility score: 19.40 -> 21.05 | Improved |
Closest papers by meaning
- Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks (CVPR 2025 · similarity 0.91)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.89)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.88)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 2025 · similarity 0.88)
- When Think-with-Image Meets Safety: What Determines Multimodal Jailbreak Robustness? (NeurIPS 2026 · similarity 0.88)
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts (AAAI 2025 · similarity 0.87)
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.
