Research map / Adversarial robustness and security / Black-box and transfer attacks
Sparse Black-Box Video Attack with Reinforcement Learning
Xingxing Wei, Huanqian Yan, Bo Hu Li
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 RL key-frame agent with R_attack reward to FGSM+NES (ablation) | SVA · un-targeted attack, 20 UCF-101 videos | MAP: 6.5037 -> 2.3723 | Improved |
| add R_rep reward to R_attack (ablation) | SVA · un-targeted attack, 20 UCF-101 videos | MAP: 2.3723 -> 2.0321 | Improved |
| add R_div reward on top of R_attack and R_rep (ablation) | SVA · un-targeted attack, 20 UCF-101 videos | MAP: 2.0321 -> 1.8624 | Improved |
| add RL agent (sparsity) vs no RL (ablation) | SVA · un-targeted attack, 20 UCF-101 videos | sparsity (%): 0.00 -> 62.35 | Improved |
| add R_rep and R_div to R_attack (ablation) | SVA · un-targeted attack, 20 UCF-101 videos | sparsity (%): 62.35 -> 74.65 | Improved |
| replace random keyframe selection with RL agent (ablation) | SVA · C3D, UCF101 20 videos | fooling rate, query, MAP | Improved |
| replace Heuristic-attack keyframe selection with proposed RL selection (SVA_opt) (ablation) | Heuristic-attack · C3D, UCF-101 and HMDB-51 | MAP, Q, S, FR | Improved |
| swap NES gradient estimator for SignSGD (ablation) | SVA · C3D, UCF101 20 videos | fooling rate: 10% better fooling rate | Mixed |
| swap NES gradient estimator for Sign-OPT (ablation) | SVA · C3D, UCF101 20 videos | fooling rate: 5% better fooling rate | Mixed |
| increase sparsity of SVAL from 50% to 70% (ablation) | SVAL · C3D UCF-101 un-targeted, 10 videos | fooling rate (%): 100.00 -> 60.0 | Hurt |
| use SVA instead of SVAL on HMDB-51 LRCN un-targeted | SVA · HMDB-51 LRCN un-targeted | MAP: 1.9479 -> 3.1570 | Hurt |
| use SVA instead of Heuristic-attack in targeted mode | SVA · UCF-101 targeted | fooling rate: - -> 32.0 | Improved |
Closest papers by meaning
- Heuristic Black-Box Adversarial Attacks on Video Recognition Models (AAAI 2020 · similarity 0.90)
- Efficient Robustness Assessment via Adversarial Spatial-Temporal Focus on Videos (TPAMI 2023 · similarity 0.90)
- Reinforcement Learning Platform for Adversarial Black-box Attacks with Custom Distortion Filters (AAAI 2025 · similarity 0.89)
- LogoStyleFool: Vitiating Video Recognition Systems via Logo Style Transfer (AAAI 2024 · similarity 0.87)
- Diffusion Guided Adversarial State Perturbations in Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- 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.
