Research map / Adversarial robustness and security / Certified and empirical robustness
Effective and Efficient Vote Attack on Capsule Networks
Jindong Gu, Baoyuan Wu, Volker Tresp
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 |
|---|---|---|---|
| Attack votes directly instead of output capsules | CapsNet with PGD · CIFAR10 | robust accuracy: 17.3{ ($\pm$1.35)} -> 4.83{ ($\pm$0.05)} | Improved |
| Attack votes directly instead of output capsules | CapsNet with PGD · SVHN | robust accuracy: 34.82{ ($\pm$2.11)} -> 8.11{ ($\pm$0.3)} | Improved |
| Add Vote-Attack examples to adversarial training | CapsNet · CIFAR10 | robust accuracy (PGD): 43.65{ ($\pm$0.85)} -> 44.12{ ($\pm$0.32)}; 45.49{ ($\pm$0.78)} -> 49.62{ ($\pm$0.56)} | Mixed |
| Add Vote-Attack examples to adversarial training | CapsNet · SVHN | robust accuracy (PGD): 30.44{ ($\pm$1.08)} -> 34.92{ ($\pm$0.98)}; 33.40{ ($\pm$1.36)} -> 39.13{ ($\pm$0.96)} | Mixed |
| Remove output-capsule routing from attack computation by attacking votes | CapsNet PGD attack · CIFAR10 test | attack time: 471.81\textit{ms} -> 414.58\textit{ms} | Improved |
| Integrate Vote-Attack into detection-aware attack | CapsNet · FMNIST | undetected rate (PGD): 78.2 -> 86.1 | Improved |
| Use Vote-Attack instead of Caps-Attack to create targeted examples | CapsNet with PGD · CIFAR10 | targeted attack success rate: 65.94{ ($\pm$0.92)} -> 75.35{ ($\pm$0.91)} | Improved |
| Generate examples by attacking votes rather than output capsules | CapsNet with PGD · CIFAR10 | transfer success rate to CapsNet: 46.43{ ($\pm$5.56)} -> 55.34{ ($\pm$6.26)} | Improved |
| Squash each vote before averaging (ablation) | Vote-Attack · CIFAR10 | robust accuracy (PGD): 4.83{ ($\pm$0.05)} -> 4.06{ ($\pm$1.12)} | No clear effect |
| Average individual-vote losses instead of loss on the squashed averaged vote (ablation) | Vote-Attack · CIFAR10 | robust accuracy (PGD): 4.83{ ($\pm$0.05)} -> 43.31{ ($\pm$2.46)} | Hurt |
| Train using Vote-Attack examples alone | CapsNet adversarial training · CIFAR10 and SVHN | robust accuracy | Hurt |
Closest papers by meaning
- Detecting and Diagnosing Adversarial Images with Class-Conditional Capsule Reconstructions (ICLR 2020 · similarity 0.88)
- Capsule Network Is Not More Robust Than Convolutional Network (CVPR 2021 · similarity 0.86)
- With Friends Like These, Who Needs Adversaries? (NeurIPS 2018 · similarity 0.84)
- Self-Routing Capsule Networks (NeurIPS 2019 · similarity 0.84)
- Countering Adversarial Images using Input Transformations (ICLR 2018 · similarity 0.84)
- Robust NAS under adversarial training: benchmark, theory, and beyond (ICLR 2024 · similarity 0.83)
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.
