Research map / Adversarial robustness and security / Black-box and transfer attacks
Scaling Laws for Black-box Adversarial Attacks
Chuan Liu, Huanran Chen, Yichi Zhang, Jun Zhu, Yinpeng Dong
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 |
|---|---|---|---|
| scale the number of surrogate models in the ensemble (log2 T from 0 to 6) with SSA-CWA (ablation) | SSA-CWA ensemble attack · Targeted attack on standard classifiers (ImageNet NIPS17) | ASR: ASR increases linearly with log T; each doubling of T gives a consistent gain | Improved |
| scale surrogate ensemble size against defended models (ablation) | SSA-CWA ensemble attack · Defended models (DiffPure, adversarially trained, HGD) | ASR: DiffPure over 30%, adversarially trained models over 20%, HGD over 50% | Improved |
| scale surrogate ensemble size, evaluated on MLLMs (ablation) | SSA-CWA ensemble attack · MLLM targets (LLaVA-1.5-7B, Qwen-2.5-VL-7B, Llama-3.2-11B-Vision) | ASR | Improved |
| scale CLIP surrogate ensemble from 1 to 12 models (ablation) | SSA-CWA on CLIP ensemble · 7 SOTA MLLMs (eps 16/255) | ASR: 85.0% on GPT-4o and 90.0% on Qwen3-VL-235B at T=12 | Improved |
| scale ensemble in the untargeted setting (ablation) | SSA-CWA ensemble attack · Untargeted attack (Swin-B, ConvNeXt-B, ResNeXt50, eps 2/255) | ASR | Improved |
| use logit-averaging ensembles (MI-FGSM, DI-MI-FGSM, VMI-FGSM) instead of gradient-aligning optimizers (ablation) | ensemble attack · Targeted sea lion, standard classifiers | ASR | Hurt |
| use advanced ensemble optimizers (MI-CWA, VMI-CWA, MI-SVRE) that resolve gradient conflict (ablation) | ensemble attack · Targeted sea lion, standard classifiers | ASR | Improved |
| add DI or VMI components to logit-ensemble MI-FGSM (ablation) | logit-averaging ensemble · Targeted sea lion, standard classifiers | ASR | No clear effect |
| increase ensemble size T for naive MI-FGSM logit ensemble (ablation) | naive logit ensemble · gradient norm analysis | gradient norm | Hurt |
Closest papers by meaning
- Scaling Trends in Language Model Robustness (ICML 2025 · similarity 0.90)
- Scaling Laws for Adversarial Attacks on Language Model Activations and Tokens (ICLR 2025 · similarity 0.89)
- Enhancing Adversarial Transferability with Checkpoints of a Single Model's Training (CVPR 2025 · similarity 0.89)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.89)
- Statistical Estimation of Adversarial Risk in Large Language Models under Best-of-N Sampling (NeurIPS 2026 · similarity 0.89)
- Adversarial Robustness Limits via Scaling-Law and Human-Alignment Studies (ICML 2024 · similarity 0.88)
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.
