Research map / Adversarial robustness and security / Certified and empirical robustness
DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks
Check 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 |
|---|---|---|---|
| Use iterative linearization to estimate minimal adversarial perturbations | DeepFool versus penalized optimization · MNIST, CIFAR-10, and ILSVRC2012 | average robustness (L2): 0.2 vs 1.0; $1.1\times 10^{-1}$ vs $3.9\times 10^{-1}$; $2.3\times 10^{-2}$ vs $1.2\times 10^{-1}$; $3.0\times 10^{-2}$ | Helped |
| Use iterative DeepFool rather than fast gradient sign | Adversarial perturbation estimation · MNIST, CIFAR-10, and ILSVRC2012 | average robustness (L2): $2.0\times 10^{-1}$ vs $2.5\times 10^{-1}$; $1.1\times 10^{-1}$ vs $1.2\times 10^{-1}$; $2.3\times 10^{-2}$ vs $2.4\time | Helped |
| Use DeepFool rather than fast gradient sign | Adversarial perturbation estimation · MNIST and CIFAR-10 | average robustness (L-infinity): 0.10 vs 0.26; 0.04 vs 0.11; 0.008 vs 0.024; 0.015 vs 0.028 | Helped |
| Fine-tune on DeepFool adversarial examples | LeNet · MNIST | adversarial robustness: improved by 50% | Helped |
| Fine-tune on DeepFool adversarial examples | NIN · CIFAR-10 | adversarial robustness: increased by about 40% | Helped |
| Fine-tune on fast gradient sign adversarial examples | LeNet and FC500-150-10 · MNIST | adversarial robustness | Hurt |
| Fine-tune on fast gradient sign adversarial examples | NIN and LeNet · CIFAR-10 | test error: 11.5% -> 21.2%; 22.6% -> 28.6% | Hurt |
| Fine-tune on DeepFool adversarial examples | LeNet · MNIST | test error: 1% -> 0.8% | Helped |
| Fine-tune on DeepFool adversarial examples | FC500-150-10 · MNIST | test error: 1.7% -> 1.5% | Helped |
| Fine-tune on DeepFool adversarial examples | NIN · CIFAR-10 | test error: 11.5% -> 11.2% | Helped |
| Fine-tune on DeepFool adversarial examples | LeNet · CIFAR-10 | test error: 22.6% -> 20.0% | Helped |
| Magnify DeepFool perturbation norms (ablation) | DeepFool adversarial examples used for fine-tuning · MNIST | adversarial robustness: 1,2,3 | Hurt |
Closest papers by meaning
- SparseFool: A Few Pixels Make a Big Difference (CVPR 2019 · similarity 0.88)
- Geometric Robustness of Deep Networks: Analysis and Improvement (CVPR 2018 · similarity 0.87)
- AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (ICLR 2020 · similarity 0.86)
- Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (ICLR 2019 · similarity 0.86)
- SuperDeepFool: a new fast and accurate minimal adversarial attack (NeurIPS 2024 · similarity 0.85)
- SURE: SUrvey REcipes for building reliable and robust deep networks (CVPR 2024 · similarity 0.85)
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.