Research map / Adversarial robustness and security / Attacks on text and graphs
DynamicPAE: Generating Scene-Aware Physical Adversarial Examples in Real-Time
Jin Hu, Xianglong Liu, Jiakai Wang, Junkai Zhang, Xianqi Yang, Haotong Qin, Yuqing Ma, Ke Xu
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 residual-guided training (reconstruction residual task) to the GDPA face-recognition patch generator (ablation) | GDPA generator · VGGFace (2% patch) | ASR / training speed: +10% ASR, 4x training speed | Improved |
| add the residual task to dynamic PAE generator training (ablation) | DynamicPAE · MSCOCO, Yolov5, Base transform | diversity of generated patches | Improved |
| increase PGD optimization steps from 8 to 512 without spatial randomness (ablation) | PGD · Inria (zero transform), Yolov5-s | AP50: 48.0% -> 1.7% | Improved |
| increase PGD EoT steps from 1 to 16 at 128 optimization steps (ablation) | PGD · MSCOCO (Base transform), Yolov5-s | AP50: 50.4% -> 29.0% | Improved |
| increase PGD optimization steps from 128 to 2048 with 1 EoT step (ablation) | PGD · MSCOCO (Base transform), Yolov5-s | AP50: 50.4% -> 27.0% | Improved |
| raise DynamicPAE attack weight lambda from 0.8 to 1 at inference (ablation) | DynamicPAE · MSCOCO, RT-DETR-L | AP50 / SSIM: AP50 0.756 -> 0.573, SSIM 0.9915 -> 0.9891 | Mixed |
| replace T-SEA with DynamicPAE as the attack method at equal training time | Yolov5-s attack · MSCOCO | AP50: 47.9% -> 14.7% | Improved |
| train DynamicPAE longer from 0.7h to 22.6h (ablation) | DynamicPAE · MSCOCO, Yolov5-s | AP50: 28.8% -> 14.7% | Improved |
| replace per-image PGD with DynamicPAE generation | PAE generation · Inria, batch of 32 | generation time: more than 2000 times speedup | Improved |
| use DynamicPAE with a robust surrogate versus a light non-robust surrogate in black-box transfer | DynamicPAE · black-box transfer | ASR: ASR +18.4% vs +6.2% | Mixed |
Closest papers by meaning
- Breaking Barriers in Physical-World Adversarial Examples: Improving Robustness and Transferability via Robust Feature (AAAI 2025 · similarity 0.86)
- MAGIC: Mastering Physical Adversarial Generation in Context Through Collaborative LLM Agents (AAAI 2026 · similarity 0.85)
- UV-Attack: Physical-World Adversarial Attacks on Person Detection via Dynamic-NeRF-based UV Mapping (ICLR 2025 · similarity 0.85)
- Embodied Active Defense: Leveraging Recurrent Feedback to Counter Adversarial Patches (ICLR 2024 · similarity 0.84)
- AdvPaint: Protecting Images from Inpainting Manipulation via Adversarial Attention Disruption (ICLR 2025 · similarity 0.84)
- PubDef: Defending Against Transfer Attacks From Public Models (ICLR 2024 · similarity 0.84)
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.
