Research map / Adversarial robustness and security / Attacks on text and graphs
DualMirage: Hunting Stealthy Multimodal LLM Agents via CAPTCHAs with Contour and Adversarial Illusions
Bei Chen, Gaolei Li, Jun Wu, Jianhua Li
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 contour illusion (no adversarial) (ablation) | DualMirage · Illusion-Selfie2Anime vs GPT-4o | ABR: 0.0 -> 78.0 | Improved |
| add adversarial illusion to contour illusion (ablation) | DualMirage · Illusion-Selfie2Anime vs GPT-4o | ASR: 0.0 -> 44.0 | Improved |
| add contour illusion to adversarial illusion (ablation) | DualMirage · Illusion-Selfie2Anime vs GPT-4o | ABR: 36.0 -> 96.0 | Improved |
| add contour illusion to adversarial illusion (ablation) | DualMirage · Illusion-MNIST l=10 vs GPT-4o | ASR: 16.0 -> 18.0 | No clear effect |
| add adversarial illusion to contour illusion (ablation) | DualMirage · Illusion-Selfie2Anime vs LLaVA-v1.5-7b | ASR: 0.0 -> 100.0 | Improved |
| apply JPEG, blur or cropping to the challenge image (ablation) | DualMirage · Robustness (avg) | ASR: 45.6 -> 15.2 (JPEG) | Hurt |
| apply JPEG compression (ablation) | DualMirage · Robustness (avg) | ABR: 90.6 -> 100.0 | Improved |
| attacker uses few-shot learning or adversarial training (ablation) | DualMirage · Robustness (avg) | ABR: 90.6 -> 90.0 (FL), 91.0 (AT) | No clear effect |
| use more attempts for humans | DualMirage · Human usability | HSR: 95.78 -> 97.48 | Improved |
| use Illusion-Selfie2Anime instead of Illusion-MNIST | DualMirage · black-box MLLMs | ASR: 9.0 -> 44.0 (GPT-4o, l=3 vs Selfie2Anime) | Improved |
Closest papers by meaning
- Open CaptchaWorld: A Comprehensive Web-based Platform for Testing and Benchmarking Multimodal LLM Agents (NeurIPS 2025 · similarity 0.85)
- Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models (ICML 2026 · similarity 0.84)
- SLADE: Shielding against Dual Exploits in Large Vision-Language Models (CVPR 2025 · similarity 0.84)
- MCA-Bench: A Multimodal Benchmark for Evaluating CAPTCHA Robustness Against VLM-based Attacks (AAAI 2026 · similarity 0.83)
- [Re] Boosting the Visual Interpretability of CLIP via Adversarial Fine-Tuning (NeurIPS 2026 · similarity 0.83)
- Spatial CAPTCHA: Generatively Benchmarking Spatial Reasoning for Human-Machine Differentiation (ICLR 2026 · similarity 0.83)
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.
