LIMANS: Linear Model of the Adversarial Noise Space

Lucas Anquetil, Jordan Frecon, Yuan Liu, Stephane Canu, Gilles Gasso

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Deep Neural Networks, Adversarial attacks, Dictionary learning

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: This work proposes to model adversarial noise as a linear combination of universal adversarial directions and bring in stochastic gradient based algorithms to learn them, leading to robust and more transferable results.

Abstract: Recent works have revealed the vulnerability of deep neural network (DNN) classifiers to adversarial attacks. Among such attacks, it is common to distinguish specific attacks adapted to each example from universal ones referred as example-agnostic. Even though specific adversarial attacks are efficient on their target DNN classifier to attack, they struggle to transfer to others. Conversely, universal adversarial attacks suffer from lower attack success. To reconcile universality and efficiency, we propose a model of the adversarial noise space allowing to frame specific adversarial perturbation as a linear combination of universal adversarial directions. We bring in two stochastic gradient based algorithms for learning these universal directions and the associated adversarial attacks. Empirical analyses conducted with the CIFAR-10 and ImageNet datasets show that LIMANS (i) enables crafting specific and robust adversarial attacks with high probability, (ii) provides a deeper understanding of DNN flaws, and (iii) shows significant ability in transferability.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4988