Research map / Adversarial robustness and security / Black-box and transfer attacks
A Direct Approach to Robust Deep Learning Using Adversarial Networks
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 |
|---|---|---|---|
| train the classifier jointly against a learned adversarial-noise generator instead of standard training | LeNet5 · MNIST, white-box FGS | accuracy: 23.70% -> 94.66% | Helped |
| train the classifier jointly against a learned adversarial-noise generator instead of standard training | ResNet-18 · SVHN, white-box FGS | accuracy: 64.64% -> 91.51% | Helped |
| train the classifier jointly against a learned adversarial-noise generator instead of standard training | ResNet · CIFAR10, white-box PGD | accuracy: 1.32% -> 44.28% | Helped |
| use adversarial PGD training instead of adversarial-network training | LeNet5 · MNIST, white-box PGD | accuracy: 87.09% -> 92.92% | Helped |
| use adversarial PGD training instead of adversarial-network training | ResNet-18 · SVHN, white-box PGD | accuracy: 37.97% -> 42.96% | Helped |
| use adversarial-network training instead of adversarial PGD training | ResNet · CIFAR10, clean | accuracy: 75.30% -> 91.08% | Helped |
| use adversarial-network training instead of adversarial PGD training | ResNet · CIFAR10, white-box FGS | accuracy: 47.63% -> 72.81% | Helped |
| use adversarial-network training instead of ensemble adversarial training | classifier · MNIST, black-box attacks | accuracy | Helped |
| use dynamic learned adversaries rather than a static set of adversaries | classifier · SVHN and CIFAR10, black-box attacks | accuracy | Mixed |
| increase the autoencoder generator width from 8 to 64 filters (ablation) | generator attacking an undefended CIFAR10 classifier · CIFAR10, attack accuracy on original model | accuracy: 31.07% -> 6.08% | Hurt |
| use an autoencoder generator with 64 filters rather than 8 filters (ablation) | adversarial-network classifier · CIFAR10, white-box PGD | accuracy: 33.56% -> 33.38% | No clear effect |
| use a random Gaussian input for the generator instead of image input (ablation) | adversarial-network classifier · CIFAR10, white-box PGD | accuracy: 33.56% -> 35.16% | No clear effect |
Closest papers by meaning
- Defense-GAN: Protecting Classifiers Against Adversarial Attacks Using Generative Models (ICLR 2018 · similarity 0.93)
- One Man's Trash Is Another Man's Treasure: Resisting Adversarial Examples by Adversarial Examples (CVPR 2020 · similarity 0.92)
- Improved Network Robustness with Adversary Critic (NeurIPS 2018 · similarity 0.91)
- Deep Defense: Training DNNs with Improved Adversarial Robustness (NeurIPS 2018 · similarity 0.91)
- Adversarial Training Can Provably Improve Robustness: Theoretical Analysis of Feature Learning Process Under Structured Data (ICLR 2025 · similarity 0.90)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · 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.