Research map / Adversarial robustness and security / Black-box and transfer attacks
PatchAttack: A Black-box Texture-based Attack with Reinforcement Learning
Chenglin Yang, Adam Kortylewski, Cihang Xie, Yinzhi Cao, Alan Yuille
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 |
|---|---|---|---|
| optimize patch color (RGB) instead of gray in monochrome patch attack (ablation) | MPA · ImageNet non-targeted, averaged over 4 networks | avg occluded area (%): 6.67 -> 5.37 | Improved |
| add texture from class-specific dictionary to patches (TPA vs MPA_RGB) (ablation) | MPA_RGB · ImageNet targeted, ResNet50 | target accuracy (%): 25.90 -> 99.70 | Improved |
| use reinforcement learning instead of Hastings sampling in the same RGB space (ablation) | HPA_RGB · ImageNet non-targeted, ResNet50 | avg occluded area (%): 16.88 -> 5.41 | Improved |
| use reinforcement learning instead of Hastings sampling in texture space (ablation) | HPA_N10_4% · ImageNet targeted, ResNet50 | avg queries: 14345 -> 8643 | Improved |
| use reinforcement learning instead of Hastings sampling in texture space (ablation) | HPA_N10_4% · ImageNet targeted, ResNet50 | target accuracy (%): 99.80 -> 99.70 | No clear effect |
| increase patch budget from 2% to 4% of image area, 10 patches (ablation) | TPA · ImageNet targeted, ResNet50 | target accuracy (%): 97.60 -> 99.70 | Improved |
| increase patch budget from 4% to 10% of image area, 10 patches (ablation) | TPA · ImageNet targeted, ResNet50 | avg queries: 8643 -> 3747 | Improved |
| use up-to-8 patches of 2% instead of up-to-4 patches of 4% (ablation) | TPA · ImageNet non-targeted, ResNet50 | avg occluded area (%): 5.06 -> 3.10 | Improved |
| add RL monochrome patch attack to Denoise-ResNet152 defense, compare to normal ResNet50 | MPA_RGB · ImageNet non-targeted | avg occluded area (%): 5.41 -> 0.48 | Improved |
| shape-biased network as a defense against TPA | TPA_N10_4% · ImageNet targeted | target accuracy (%): 99.70 -> 100.00 | No clear effect |
| apply Local Gradients Smoothing defense | TPA_N10_4% · ImageNet targeted, ResNet50 | target accuracy (%): 99.70 -> 97.50 | No clear effect |
| use white-box adversarial patch dictionary from VGG19 on other networks | AdvPatchDict · ImageNet targeted | target accuracy (%): 98.20 -> 0.00 | Hurt |
Closest papers by meaning
- Revisiting Adversarial Patch Defenses on Object Detectors: Unified Evaluation, Large-Scale Dataset, and New Insights (ICCV 2025 · similarity 0.88)
- Towards Practical Certifiable Patch Defense With Vision Transformer (CVPR 2022 · similarity 0.87)
- Efficient Decision-based Black-box Patch Attacks on Video Recognition (ICCV 2023 · similarity 0.87)
- PAD: Patch-Agnostic Defense against Adversarial Patch Attacks (CVPR 2024 · similarity 0.86)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.86)
- Amnesia as a Catalyst for Enhancing Black Box Pixel Attacks in Image Classification and Object Detection (NeurIPS 2024 · similarity 0.85)
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.
