Research map / Adversarial robustness and security / Attacks on text and graphs
Detecting and Diagnosing Adversarial Images with Class-Conditional Capsule Reconstructions
Yao Qin, Nicholas Frosst, Sara Sabour, Colin Raffel, Garrison Cottrell, Geoffrey Hinton
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 CapsNet class-conditional reconstructions instead of CNN+CR | Adversarial-image detection · MNIST, white-box targeted PGD | undetected rate: 0 -> 0 | No clear effect |
| Use CapsNet rather than CNN+CR | Adversarial-image detection · Fashion-MNIST and SVHN, targeted and untargeted R-PGD | undetected rate: 84.4 -> 29.8; 86.0 -> 75.5; 97.9 -> 79.2; 99.5 -> 97.5 | Improved |
| Add class-conditional masking to CNN reconstruction (CNN+CR instead of CNN+R) (ablation) | CNN adversarial-image detector · White-box targeted R-PGD | undetected rate: 71.2 -> 68.1; 88.4 -> 84.4; 99.3 -> 97.9 | Improved |
| Remove class-conditional masking and reconstruct using all capsules (CapsNet All) (ablation) | CapsNet detector · CIFAR-10, white-box R-PGD | undetected rate | Hurt |
| Replace capsule masking with the winning-capsule vector as reconstruction input (DeepCaps) (ablation) | CapsNet detector · CIFAR-10, white-box R-PGD | undetected rate | Hurt |
| Use two-stage optimization instead of a single combined loss for targeted reconstructive attacks (ablation) | Reconstructive attack · Targeted reconstructive attacks | attack success rate | Improved |
| Add reconstruction-loss minimization to a standard PGD attack (R-PGD) | White-box adversarial attack · MNIST, Fashion-MNIST, and SVHN | attack success rate | Hurt |
Closest papers by meaning
- Adversarial Attacks Are Reversible With Natural Supervision (ICCV 2021 · similarity 0.89)
- One Man's Trash Is Another Man's Treasure: Resisting Adversarial Examples by Adversarial Examples (CVPR 2020 · similarity 0.89)
- Effective and Efficient Vote Attack on Capsule Networks (ICLR 2021 · similarity 0.88)
- PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples (ICLR 2018 · similarity 0.88)
- On Adversarial Training without Perturbing all Examples (ICLR 2024 · similarity 0.88)
- Wide Two-Layer Networks can Learn from Adversarial Perturbations (NeurIPS 2024 · similarity 0.87)
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.
