Universal Unlearnable Examples: Cluster-wise Perturbations without Label-consistency

Jiaming Zhang

Keywords: pravicy-preserving, unlearnable example, adversarial attack, poisoning attack

Abstract: There is a growing interest in employing unlearnable examples against privacy leaks on the Internet, which prevents the unauthorized models from being properly trained by adding invisible image noise. However, existing attack methods rely on an ideal assumption called label-consistency. In this work, we clarify a more practical scenario called \emph{label-inconsistency} that allows hackers and protectors to hold different labels for the same image. Inspired by disrupting the \emph{uniformity} and \emph{discrepancy}, we present a novel method called \emph{UniversalCP} on label-inconsistency scenario, which generates the universal unlearnable examples by cluster-wise perturbation. Furthermore, we investigate a new strategy for selecting the CLIP as the surrogate model, since vision-and-language pre-training models are trained on large-scale data and more semantic supervised information. We also verify the effectiveness of the proposed methods and the strategy for selecting surrogate models under a variety of experimental settings including black-box backbones, datasets and even commercial platforms Microsoft {\tt Azure} and Baidu {\tt PaddlePaddle}.

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.

TL;DR: We proposed a novel method called UniversalCP, which is effective in a more practical scenario.

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

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning