Research map / Adversarial robustness and security / Attacks on text and graphs
Defense-GAN: Protecting Classifiers Against Adversarial Attacks Using Generative Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Defense-GAN projection (GD reconstruction, L=200, R=10) before classifier | MNIST classifiers under FGSM white-box attack · MNIST, FGSM white-box eps=0.3 | accuracy: Classifier A 0.217 -> 0.988 | Helped |
| add Defense-GAN projection before classifier | MNIST classifiers under FGSM black-box attack · MNIST, black-box FGSM eps=0.3 | accuracy: Classifier A 0.6343 -> 0.9312 | Helped |
| use Defense-GAN instead of MagNet reformer | MNIST black-box FGSM · MNIST, black-box FGSM eps=0.3 | accuracy: Classifier A 0.6937 (MagNet) -> 0.9312 | Helped |
| adversarial training (FGSM) vs Defense-GAN under white-box CW | MNIST classifiers · MNIST, CW l2 white-box | accuracy: Adv. Tr. 0.077 vs Defense-GAN-Rec 0.989 for Classifier A | Hurt |
| adversarial training with eps=0.3 under black-box FGSM | Classifier B on MNIST · MNIST, black-box FGSM eps=0.3 | accuracy: Adv. Tr. 0.2092 vs Defense-GAN-Rec 0.9057 for Classifier B/Substitute B | Hurt |
| increase GD iterations L from 25 to 400 (ablation) | Defense-GAN-Rec · MNIST, black-box FGSM eps=0.3, Model F | accuracy under attack: 0.7955 (L=25) -> 0.9332 (L=400) | Helped |
| increase GD iterations L beyond 400 (to 800) under black-box attack (ablation) | Defense-GAN-Rec · MNIST, black-box FGSM eps=0.3, Model F | accuracy under attack: 0.9332 (L=400) -> 0.9319 (L=800) -> 0.9081 (L=1600) | Mixed |
| increase random restarts R from 1 to 20 (ablation) | Defense-GAN-Rec (L=100) · MNIST, black-box FGSM eps=0.3, Model F | accuracy under attack: 0.5329 (R=1) -> 0.9456 (R=20) | Helped |
| increase GD iterations L (25 vs 200) (ablation) | Defense-GAN under white-box FGSM · MNIST, white-box FGSM eps=0.3, Classifier A | accuracy: 0.947 (L=25) vs 0.988 (L=200) for Classifier A | Helped |
| use Defense-GAN on CelebA gender classification | Classifier A on CelebA under FGSM white-box · CelebA gender, FGSM eps=0.3 | accuracy: 0.0870 (no defense) -> 0.9255 (Defense-GAN-Rec) | Helped |
| Defense-GAN on Fashion-MNIST black-box attack | Classifier A on F-MNIST · F-MNIST, black-box FGSM eps=0.3 | accuracy: 0.5131 (no defense) -> 0.586 (Defense-GAN-Rec) | Helped |
Closest papers by meaning
- A Direct Approach to Robust Deep Learning Using Adversarial Networks (ICLR 2019 · similarity 0.93)
- Deep Defense: Training DNNs with Improved Adversarial Robustness (NeurIPS 2018 · similarity 0.91)
- PixelDefend: Leveraging Generative Models to Understand and Defend against Adversarial Examples (ICLR 2018 · similarity 0.90)
- One Man's Trash Is Another Man's Treasure: Resisting Adversarial Examples by Adversarial Examples (CVPR 2020 · similarity 0.89)
- Constructing Unrestricted Adversarial Examples with Generative Models (NeurIPS 2018 · similarity 0.89)
- Rob-GAN: Generator, Discriminator, and Adversarial Attacker (CVPR 2019 · similarity 0.89)
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.