Research map / Adversarial robustness and security / Attacks on text and graphs
FORCE: Transferable Visual Jailbreaking Attacks via Feature Over-Reliance CorrEction
Runqi Lin, Alasdair Paren, Suqin Yuan, Muyang Li, Philip Torr, Adel Bibi, Tongliang Liu
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 spectral rescaling alone to PGD (ablation) | PGD visual jailbreak · MaliciousInstruct, Idefics3-8B-Llama3 | ASR (%): 53.00 -> 59.00 | Improved |
| add layer-aware feature regularisation alone to PGD (ablation) | PGD visual jailbreak · MaliciousInstruct, Idefics3-8B-Llama3 | ASR (%): 53.00 -> 55.00 | Improved |
| combine layer regularisation with spectral rescaling (ablation) | FORCE · MaliciousInstruct, Idefics3-8B-Llama3 | ASR (%): 53.00 -> 64.00 | Improved |
| apply FORCE instead of PGD against adapter-based MLLMs | PGD visual jailbreak · MaliciousInstruct, Llava-v1.6-mistral-7b | ASR (%): 61.00 -> 69.00 | Improved |
| apply FORCE instead of PGD against early-fusion MLLMs | PGD visual jailbreak · MaliciousInstruct, Qwen2.5-VL-7B-Instruct | ASR (%): 5.00 -> 11.00 | Improved |
| apply FORCE instead of PGD against Gemini-2.5-Pro | PGD visual jailbreak · MaliciousInstruct | ASR (%): 10.00 -> 10.00 | No clear effect |
| apply FORCE instead of PGD against Claude-Sonnet-4 | PGD visual jailbreak · AdvBench | ASR (%): 1.00 -> 1.00 | No clear effect |
| apply FORCE in the zero-shot single-query setting | PGD visual jailbreak · MaliciousInstruct, Llava-v1.6-mistral-7b | ASR (%): 26.00 -> 26.00 | No clear effect |
| use blank initial image with FORCE (ablation) | FORCE · MaliciousInstruct, Idefics3-8B-Llama3 | ASR (%): 64.00 -> 83.00 | Improved |
| ensemble optimisation over two source models | PGD visual jailbreak · MaliciousInstruct, Llava-v1.6 | ASR (%): 61.00 -> 59.00 | Hurt |
| add MI-FGSM momentum to PGD | PGD visual jailbreak · MaliciousInstruct, Llava-v1.6 | ASR (%): 61.00 -> 49.00 | Hurt |
| add uniform or Gaussian noise defence to FORCE attacks | FORCE · MaliciousInstruct, Idefics3-8B-Llama3 | ASR (%): 64.00 -> 62.00 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports FORCE as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Dissecting the Safety Circuit: Neuronal Intervention for Transferable Adversarial Attacks on VLMs (ICML 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
- Distraction is All You Need for Multimodal Large Language Model Jailbreaking (CVPR 2025 · similarity 0.91)
- MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs (ICLR 2026 · similarity 0.91)
- Failures to Find Transferable Image Jailbreaks Between Vision-Language Models (ICLR 2025 · similarity 0.90)
- MMJ-Bench: A Comprehensive Study on Jailbreak Attacks and Defenses for Vision Language Models (AAAI 2025 · similarity 0.90)
- Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models (ICLR 2024 · similarity 0.90)
- Understanding and Enhancing the Transferability of Jailbreaking Attacks (ICLR 2025 · similarity 0.90)
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.
