Research map / Adversarial robustness and security / Black-box and transfer attacks
Adversarial Defense by Restricting the Hidden Space of Deep Neural Networks
Aamir Mustafa, Salman Khan, Munawar Hayat, Roland Goecke, Jianbing Shen, Ling Shao
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 |
|---|---|---|---|
| add prototype conformity loss to softmax training | ResNet-110 · CIFAR-10 white-box PGD (eps 0.03) | robust accuracy: 0.01 -> 27.2 | Improved |
| add prototype conformity loss to softmax training | ResNet-110 · CIFAR-100 white-box PGD (eps 0.03) | robust accuracy: 0.17 -> 25.9 | Improved |
| add prototype conformity loss to softmax training | CNN-6 · MNIST white-box PGD (eps 0.3) | robust accuracy: 0.0 -> 19.9 | Improved |
| add prototype conformity loss to softmax training | ResNet-110 · CIFAR-10 black-box PGD (eps 0.03, VGG-19 source) | robust accuracy: 29.1 -> 76.4 | Improved |
| add prototype conformity loss | ResNet-110 · CIFAR-10 clean | accuracy: 90.80 -> 90.45 | No clear effect |
| add PGD adversarial training to PC loss | ResNet-110 · CIFAR-10 white-box PGD (eps 0.03) | robust accuracy: 27.2 -> 46.7 | Improved |
| use PGD instead of FGSM adversarial examples for AdvTrain | ResNet-110 · CIFAR-10 white-box PGD (eps 0.03) | robust accuracy: 42.5 -> 46.7 | Improved |
| add adversarial training to PC loss | ResNet-110 · CIFAR-100 clean | accuracy: 71.90 -> 69.11 | Hurt |
| attack the joint PC+CE loss (adaptive white-box) | PC loss model, ResNet-110 · CIFAR-10 PGD (eps 8/255) | robust accuracy: 27.2 -> 27.09 | No clear effect |
| apply L_PC after Layer 5 (ablation) | ResNet-110 · CIFAR-10 PGD (eps 0.03), L_PC layer ablation | robust accuracy: 0.01 -> 13.56 | Improved |
| apply L_PC after Layers 4 and 5 instead of Layer 5 only (ablation) | ResNet-110 · CIFAR-10 PGD (eps 0.03), L_PC layer ablation | robust accuracy: 13.56 -> 20.70 | Improved |
| apply L_PC after Layers 3, 4 and 5 instead of 4 and 5 (ablation) | ResNet-110 · CIFAR-10 PGD (eps 0.03), L_PC layer ablation | robust accuracy: 20.70 -> 27.23 | Improved |
Closest papers by meaning
- Deeply Supervised Discriminative Learning for Adversarial Defense (TPAMI 2020 · similarity 0.97)
- Class-Disentanglement and Applications in Adversarial Detection and Defense (NeurIPS 2021 · similarity 0.89)
- Stylized Adversarial Defense (TPAMI 2022 · similarity 0.89)
- With Friends Like These, Who Needs Adversaries? (NeurIPS 2018 · similarity 0.89)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.88)
- Feature Space Perturbations Yield More Transferable Adversarial Examples (CVPR 2019 · similarity 0.88)
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.
