Research map / Adversarial robustness and security / Attacks on text and graphs
DART: Distribution-Aware Adaptive Relational Transfer for Adversarial Attacks against Closed-Source MLLMs
Kaidi Hu, Guancheng Wan, Xiao Luo, Ruigang 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 Energy Distance distribution alignment with M=5 augmentations (vs point-wise M=1) (ablation) | DART · GPT-4o ablation | ASR: 83.0 -> 89.5 | Improved |
| add intra-distribution terms (Energy Distance vs AvgCosine/EoT) (ablation) | DART · GPT-4o ablation | ASR: 82.5 -> 89.5 | Improved |
| add data augmentation with simple gradient smoothing only (AvgCosine, M=5) vs point-wise (ablation) | point-wise attack · GPT-4o ablation | ASR: 83.0 -> 82.5 | No clear effect |
| use Energy Distance instead of MMD (ablation) | DART · GPT-4o | ASR | Improved |
| increase GNN router ensemble size from K=1 to K=2 (ablation) | DART · GPT-4o | ASR | Improved |
| increase ensemble size from K=2 to K=3 (ablation) | DART · GPT-4o | ASR | Hurt |
| increase augmented samples M from 2 to 5 (ablation) | DART · GPT-4o | ASR: +3.0% from M=1 to M=2; saturates beyond M=5 | Improved |
| use heterogeneous global/local weighting eta=[1.8, 2.1, 4.8] instead of uniform (ablation) | DART · GPT-4o | ASR | Improved |
| replace FOA-Attack with DART on open-source Qwen2.5-VL-7B | DART · Open-source MLLMs | ASR: 80.0 -> 86.5 | Improved |
| replace FOA-Attack with DART on Claude-3.7 | DART · Closed-source MLLMs | ASR: 23.5 -> 32.0 | Improved |
| apply Gaussian blur defense (r=1.0) to adversarial images | DART vs FOA-Attack · GPT-4o | ASR: 89.5 -> 65.0 (FOA 85.5 -> 53.0) | Hurt |
| reduce M from 5 to 2 for efficiency (ablation) | DART · GPT-4o | ASR: 89.5 -> 86.0; time 175 s -> 86 s | Mixed |
Closest papers by meaning
- Dual Feature-Relational Alignment for Transferable Targeted Attacks on MLLMs (NeurIPS 2026 · similarity 0.88)
- TRS: Transferability Reduced Ensemble via Promoting Gradient Diversity and Model Smoothness (NeurIPS 2021 · similarity 0.88)
- Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment (NeurIPS 2025 · similarity 0.88)
- A Frustratingly Simple Yet Highly Effective Attack Baseline: Over 90% Success Rate Against the Strong Black-box Models of GPT-4.5/4o/o1 (NeurIPS 2025 · similarity 0.87)
- PubDef: Defending Against Transfer Attacks From Public Models (ICLR 2024 · similarity 0.87)
- Enhancing Adversarial Transferability with Adversarial Weight Tuning (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.
