Constructing Semantics-Aware Adversarial Examples with Probabilistic Perspective

Andi Zhang, Damon Wischik

Primary Area: generative models

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: adversarial examples, probabilistic generative models, energy-based models

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

Abstract: In this study, we introduce a novel, probabilistic perspective for generating adversarial examples. Within this view, geometric constraints on adversarial examples are interpreted as distributions, facilitating the transition from geometric constraints to data-driven semantic constraints. Proceeding from this perspective, we develop an innovative approach for generating semantics-aware adversarial examples in a principled manner. Our approach empowers individuals to incorporate their personal comprehension of semantics into the model. Through human evaluation, we validate that our semantics-aware adversarial examples maintain their inherent meaning. Experimental findings on the MNIST, SVHN and CIFAR10 datasets demonstrate that our semantics-aware adversarial examples can effectively circumvent robust adversarial training methods tailored for traditional adversarial attacks.

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: 4207