MearcAll papers

arXiv 1706.06083 · stat.ML cs.LG cs.NE

Towards Deep Learning Models Resistant to Adversarial Attacks

Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, Adrian Vladu

Claims were extracted automatically by reducer-1 on 2026-11-12. The paper's license does not allow republishing it, so long claims are shortened to a quotation. Read the full text on arXiv.

No reproduction recordedres_e6332734
Our best MNIST model achieves an accuracy of more than 89% against the strongest adversaries in our test suite.

Introduction · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_917fe0b3
Our CIFAR10 model achieves an accuracy of 46% against 2 the same adversary.

Introduction · page 2 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_a993c82e
Furthermore, in case of the weaker black box/transfer attacks, our MNIST and CIFAR10 networks achieve the accuracy of more than 95% and 64%, respectively.

Introduction · page 3 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_30e9cce6
We argue that the loss landscape corresponding to this problem has a surprisingly tractable structure of local maxima.

Towards Universally Robust Networks · page 5 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_35432e38
Surprisingly, our experiments show that the inner problemis tractable after all, at least from the perspective of first-order methods.

The Landscape of Adversarial Examples · page 5 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_af43039c
Our experiments show that the local maxima found by PGD all have similar loss values, both for normally trained networks and adversarially trained networks.

First-Order Adversaries · page 7 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_5eb54408
However, our experiments suggest that such better local maxima are hard to find with first order methods: even a large number of random restarts did not find function values with … (excerpt)

First-Order Adversaries · page 7 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_9b463190
Despite the fact that the exact assumptions of Danskin’s theorem do not hold for our problem (the function is not continuously differentiable due to ReLU and max-pooling units, and we … (excerpt)

Descent Directions for Adversarial Training · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_4a8b9a26
We need to also argue that the value of the problem (i.e. the final loss we 8 achieve against adversarial examples) is small, thus providing guarantees for the performance of … (excerpt)

Network Capacity and Adversarial Robustness · page 8 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_87218bfa
This network can achieve an accuracy of 95.2% when trained with natural examples.

Network Capacity and Adversarial Robustness · page 9 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedres_65139fdf
These plots illustrate that we can consistently reduce the value of the inner problem of the saddle point formulation (2.1), thus producing an increasingly robust classifier. trained with natural examples, … (excerpt)

Experiments: Adversarially Robust Deep Learning Models · page 12 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_c837fc79
For the MNIST dataset, our networks are very robust, achieving high accuracy for a wide range of powerful 𝓁∞-bound adversaries and large perturbations.

Conclusion · page 15 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_c1c8720f
Our experiments on CIFAR10 have not reached the same level of performance yet.

Conclusion · page 15 · Automated extraction, not reviewed · 0 reviewed reproductions

No reproduction recordedclm_c951fc87
However, our results already show that our techniques lead to significant increase in the robustness of the network.

Conclusion · page 15 · Automated extraction, not reviewed · 0 reviewed reproductions

Reproduced one of these claims?

Record your result in the Mearc app on your own machine, then use Publish this result with this paper's id (1706.06083) and the claim id shown next to the claim. Mearc creates a file that holds only what you choose to publish. Add it to the public claim data by pull request; a maintainer other than you reviews it before it counts.