Research map / Adversarial robustness and security / Attacks on text and graphs
Catastrophic Child's Play: Easy to Perform, Hard to Defend Adversarial Attacks
Chih-Hui Ho, Brandon Leung, Erik Sandstrom, Yen Chang, Nuno Vasconcelos
Check 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 pose variation rather than camera shake perturbations | ImageNet-trained CNNs · ImageNet | recognition rate: 73.7 -> 47.2 | Hurt |
| Add frontal-pose real-world examples to defense training | CNNs under camera shake attacks · ImageNet to Frontal | recognition rate: 73.7 -> 82.0 | Improved |
| Add frontal-pose real-world examples to defense training | CNNs under pose variation attacks · ImageNet to Frontal | recognition rate: 47.2 -> 63.7 | Improved |
| Use all camera-shake and pose-variation defense images instead of frontal-only images | CNNs under camera shake attacks · Frontal to All | recognition rate: 82.0 -> 87.1 | Improved |
| Use all camera-shake and pose-variation defense images instead of frontal-only images | CNNs under pose variation attacks · Frontal to All | recognition rate: 63.7 -> 79.1 | Improved |
| Apply Color Jitter defense | CNNs under camera shake attacks · Avg | recognition rate: 80.9 -> 82.7 | No clear effect |
| Use gradient-based defenses instead of no defense | CNNs under camera shake attacks · Avg | recognition rate: 80.9 -> 80.0 | Hurt |
| Use gradient-based defenses instead of no defense | CNNs under pose variation attacks · Avg | recognition rate: 63.3 -> 59.3 | Hurt |
| Craft attacks by selecting the most adversarial instance from K=5 candidates rather than sampling randomly | CNNs under camera shake attacks · ImageNet | recognition rate: 73.7 -> 51.3 | Hurt |
| Craft attacks by selecting the most adversarial instance from K=5 candidates rather than sampling randomly | CNNs under pose variation attacks · ImageNet | recognition rate: 47.2 -> 33.0 | Hurt |
| Use camera shake rather than pose variation as an object-recognition perturbation | Turk semantic perceptibility assessment · Object recognition Turk experiment | SIPR (%): 81.5 -> 91.6 | Improved |
Closest papers by meaning
- Generative Adversarial Perturbations (CVPR 2018 · similarity 0.87)
- Attack to Explain Deep Representation (CVPR 2020 · similarity 0.87)
- Physical Adversarial Attack Meets Computer Vision: A Decade Survey (TPAMI 2024 · similarity 0.87)
- Adversarial camera stickers: A physical camera-based attack on deep learning systems (ICML 2019 · similarity 0.87)
- PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples (ICLR 2018 · similarity 0.87)
- Merry Go Round: Rotate a Frame and Fool a DNN (CVPR 2022 · similarity 0.86)
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.
