Research map / Adversarial robustness and security / Certified and empirical robustness
Scaling up the Randomized Gradient-Free Adversarial Attack Reveals Overestimation of Robustness Using Established Attacks
Francesco Croce, Jonas Rauber, Matthias Hein
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 the randomized gradient-free attack instead of PGD-10k | PGD-10k · MNIST | average difference to best l2 robust accuracy: 0.1011 -> 0.0051 | Improved |
| use the randomized gradient-free attack instead of PGD-10k | PGD-10k · MNIST | maximum difference to best l2 robust accuracy: 0.5040 -> 0.0500 | Improved |
| use the randomized gradient-free attack instead of CW-100k | CW-100k · GTS | average difference to best l2 robust accuracy: 0.0172 -> 0 | Improved |
| use the randomized gradient-free attack instead of CW-100k | CW-100k · CIFAR-10 | average difference to best l2 robust accuracy: 0.0037 -> 0.0060 | Hurt |
| use the randomized gradient-free attack instead of CW-100k | CW-100k · CIFAR-10 | maximum difference to best l2 robust accuracy: 0.0180 -> 0.0140 | Improved |
| use 10000 random restarts instead of a single start for PGD | PGD-1 · MNIST | average difference to best l2 robust accuracy: 0.2367 -> 0.1011 | Improved |
| increase CW iterations from 10k to 100k | CW-10k · MNIST | average difference to best l2 robust accuracy: 0.1701 -> 0.1681 | No clear effect |
| use the randomized gradient-free attack instead of PGD-10k on an l_inf-adversarially trained model | PGD-10k · MNIST l_inf-at, eps=2.0 | l2 robust accuracy upper bound: 0.536 -> 0.032 | Improved |
| use the randomized gradient-free attack instead of DeepFool | DeepFool · MNIST l_inf-at, eps=2.0 | l2 robust accuracy upper bound: 0.918 -> 0.032 | Improved |
| use the randomized gradient-free attack instead of PGD-1k on provably robust l_inf-MMR model | PGD-1k · MNIST l_inf-MMR-at, eps=2.0 | l2 robust accuracy upper bound: 0.310 -> 0.048 | Improved |
| use the randomized gradient-free attack instead of PGD-1k on l_inf-KW model | PGD-1k · MNIST l_inf-KW, eps=2.0 | l2 robust accuracy upper bound: 0.226 -> 0.148 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DeepFool — DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks (CVPR 2016) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Accurate, reliable and fast robustness evaluation (NeurIPS 2019 · similarity 0.94)
- Theoretical evidence for adversarial robustness through randomization (NeurIPS 2019 · similarity 0.92)
- Towards Deep Learning Models Resistant to Adversarial Attacks (ICLR 2018 · similarity 0.92)
- Certified Adversarial Robustness with Additive Noise (NeurIPS 2019 · similarity 0.91)
- Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks (ICML 2020 · similarity 0.91)
- SuperDeepFool: a new fast and accurate minimal adversarial attack (NeurIPS 2024 · similarity 0.90)
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.
