Research map / Adversarial robustness and security / Black-box and transfer attacks
FeatureFool: Zero-Query Fooling of Video Models via Feature Map
Duoxun Tang, Xi Xiao, Guangwu Hu, Kangkang Sun, Xiao Yang, Dongyang Chen, Qing Li, Yong-jie Yin, Jiyao Wang
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 max-optical-flow frame instead of a random frame for the feature map (ablation) | FeatureFool (C3D) · UCF-101 | ASR: 53% -> 70% | Improved |
| use max-optical-flow frame instead of a random frame for the feature map (ablation) | FeatureFool (I3D) · UCF-101 | ASR: 57% -> 74% | Improved |
| single max-flow frame instead of computing feature map on every frame (ablation) | FeatureFool (C3D) · UCF-101 | ASR: 65% -> 70% | Improved |
| single max-flow frame instead of computing feature map on every frame (ablation) | FeatureFool (I3D) · UCF-101 | ASR: 70% -> 74% | Improved |
| select attack video by SSIM or cosine similarity instead of random (ablation) | FeatureFool · UCF-101/HMDB-51/Kinetics-400 | ASR: about 70% in all cases | No clear effect |
| use feature map from a different source architecture (ablation) | FeatureFool · UCF-101 | ASR: 70.0% -> 66.0% (C3D source to I3D victim) | Hurt |
| use feature map from a different source dataset (ablation) | FeatureFool · UCF-101 | ASR: 70.0% -> 68.0% (C3D-UCF-101 source to HMDB-51 victim) | Hurt |
| increase injection strength alpha from 0.1 to 1.0 (ablation) | FeatureFool (C3D) · HMDB-51 | ASR / SSIM / PSNR | Mixed |
| use raw Guided Back-propagation feature map instead of FullGrad or Grad-Cam (ablation) | FeatureFool · UCF-101 | ASR and quality | Improved |
| replace query-based attack with zero-query feature map attack | FeatureFool vs BSC · Kinetics-400 (C3D) | ASR: 76% -> 70% | Hurt |
| replace baselines with FeatureFool on Video-LLMs | FeatureFool vs BSC · Video-LLM harmful clips | ASR: 35.00% -> 76.25% | Improved |
| apply FeatureFool under TS defense | FeatureFool vs PatchAttack · UCF-101 (C3D) | Residual-ASR: 55.0% -> 66.0% | Improved |
Closest papers by meaning
- Heuristic Black-Box Adversarial Attacks on Video Recognition Models (AAAI 2020 · similarity 0.85)
- DifAttack: Query-Efficient Black-Box Adversarial Attack via Disentangled Feature Space (AAAI 2024 · similarity 0.83)
- Pixel2Feature Attack (P2FA): Rethinking the Perturbed Space to Enhance Adversarial Transferability (ICML 2025 · similarity 0.83)
- LogoStyleFool: Vitiating Video Recognition Systems via Logo Style Transfer (AAAI 2024 · similarity 0.83)
- Amnesia as a Catalyst for Enhancing Black Box Pixel Attacks in Image Classification and Object Detection (NeurIPS 2024 · similarity 0.82)
- BRUSLEATTACK: A QUERY-EFFICIENT SCORE- BASED BLACK-BOX SPARSE ADVERSARIAL ATTACK (ICLR 2024 · similarity 0.82)
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.
