Research map / Adversarial robustness and security / Black-box and transfer attacks
Dissecting the Safety Circuit: Neuronal Intervention for Transferable Adversarial Attacks on VLMs
Chunlong Xie, Kangjie Chen, Shangwei Guo, Shudong Zhang, Jiamou Liu, Tianwei Zhang, Tao Xiang
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 |
|---|---|---|---|
| remove defensive circuit suppression loss (ablation) | SCIA · Qwen-2.5-VL target | ASR: 0.487 -> 0.651 | Improved |
| add transferable circuit amplification (ablation) | SCIA · Qwen-2.5-VL target | ASR: 0.558 -> 0.651 | Improved |
| add contrastive semantic steering loss (ablation) | SCIA · InstructBlip target | ASR: 0.378 -> 0.517 | Improved |
| add spectral smoothness regularization (ablation) | SCIA · Qwen-2.5-VL target | ASR: 0.598 -> 0.651 | Improved |
| increase perturbation budget from 4/255 to 16/255 (ablation) | SCIA · 3 victim VLMs average | ASR: 0.220 -> 0.546 | Improved |
| increase perturbation budget from 16/255 to 32/255 (ablation) | SCIA · 3 victim VLMs average | ASR: 0.546 -> 0.569 | No clear effect |
| increase circuit size from 0.5% to 5% of layer neurons (ablation) | SCIA · Average over victims | ASR: 0.462 -> 0.334 | Hurt |
| use very small circuit (0.05% of neurons) (ablation) | SCIA · Average over victims | ASR: 0.363 vs 0.462 at 0.5% | Hurt |
| deactivate identified defensive circuit neurons (ablation) | LLaVA-v1.6 · AdvBench | ASR: 0.24 -> 0.86 (random 0.26) | Improved |
| deactivate identified transferable circuit neurons (ablation) | surrogate-generated adversarial examples · LLaVA-v1.6 victim | ASR: 0.79 -> 0.33 (random 0.74) | Hurt |
| apply Gaussian smoothing as input preprocessing defense | SCIA attack · 3 black-box victims | ASR: 0.546 -> 0.319 | Hurt |
| apply bit-depth reduction as defense | SCIA attack · 3 black-box victims | ASR: 0.546 -> 0.473 | No clear effect |
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.
- FORCE — FORCE: Transferable Visual Jailbreaking Attacks via Feature Over-Reliance CorrEction (CVPR 2026) · better on 1 of 1 shared result
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
- Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection (CVPR 2026 · similarity 0.90)
- ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time (ICLR 2025 · similarity 0.88)
- Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models (ICML 2024 · similarity 0.88)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility (ICML 2026 · similarity 0.87)
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications (ICML 2024 · 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.
