Research map / Adversarial robustness and security / Black-box and transfer attacks
Efficient Decision-based Black-box Patch Attacks on Video Recognition
Kaixun Jiang, Zhaoyu Chen, Hao Huang, Jiafeng Wang, Dingkang Yang, Bo Li, Yan Wang, Wenqiang Zhang
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 |
|---|---|---|---|
| use evolutionary STDE instead of reinforcement-learning TPA in decision-based setting | STDE · UCF-101 untargeted (C3D) | fooling rate: 56.40 -> 100 | Improved |
| use STDE instead of score-based BSCA while using only hard labels | STDE · UCF-101 untargeted (C3D) | fooling rate: 70.90 (BSCA*) -> 100 | Improved |
| port STDE to bullet-screen patches (STDE-W) versus BSCA | BSCA · UCF-101 untargeted (C3D) | fooling rate / queries: 59.4 -> 65.3; AQN 4821 -> 1229 | Improved |
| fill patch texture with target video instead of white (ablation) | STDE bullet screen · UCF-101 targeted (C3D) | fooling rate: 17.8 -> 23.8 | Improved |
| add temporal difference keyframe selection (ablation) | STDE · UCF-101 targeted (C3D) | AOA: 24.90 -> 18.70 | Improved |
| add temporal difference keyframe selection (ablation) | STDE · UCF-101 untargeted (C3D) | AOA: 4.19 -> 4.06 | Improved |
| add crossover operation (ablation) | STDE · UCF-101 targeted (C3D) | AOA: AOA increases by 20% without crossover | Improved |
| use l0 norm instead of l2 norm in fitness function (ablation) | STDE · UCF-101 (C3D) | AOA: AOA drops by 1% targeted, 0.16% untargeted | Improved |
| use target video texture instead of Gaussian noise or monochrome blocks (ablation) | STDE · UCF-101 untargeted (C3D) | AOA: 5% increase in AOA with alternatives | Improved |
| choose different source/target videos (ablation) | STDE · UCF-101 (C3D, NL, TPN) | fooling rate: 100 on all models | No clear effect |
| apply patch defenses LGS and DW to input | STDE · NL, UCF-101 | fooling rate: 100 -> 98 (DW untargeted) | No clear effect |
| use STDE instead of BSCA for image quality | BSCA · UCF-101 (C3D) | PSNR: 22.74 -> 34.04 | Improved |
Closest papers by meaning
- Attacking Video Recognition Models with Bullet-Screen Comments (AAAI 2022 · similarity 0.87)
- PatchAttack: A Black-box Texture-based Attack with Reinforcement Learning (ECCV 2020 · similarity 0.87)
- Decision-based Black-box Attack Against Vision Transformers via Patch-wise Adversarial Removal (NeurIPS 2022 · similarity 0.86)
- Adversarial Attacks on Black Box Video Classifiers: Leveraging the Power of Geometric Transformations (NeurIPS 2021 · similarity 0.86)
- A Unified, Resilient, and Explainable Adversarial Patch Detector (CVPR 2025 · similarity 0.85)
- LogoStyleFool: Vitiating Video Recognition Systems via Logo Style Transfer (AAAI 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.
