Availability Attacks Need to Create Shortcuts for Contrastive Learning

Yihan Wang, Yifan Zhu, Xiao-Shan Gao

Primary Area: societal considerations including fairness, safety, privacy

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: Availability attacks, supervised unlearnability, contrastive unlearnability, augmented unlearnable example attack, augmented adversarial poisoning attack, transferable unlearnability, worst-case unlearnability

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

TL;DR: We propose the augmented unlearnable example (AUE) attack and augmented adversarial poisoning (AAP) attack which largely boost the worst-case unlearnability across multiple supervised and contrastive algorithms.

Abstract:

Availability attacks can prevent the unauthorized use of private data and commercial datasets by generating imperceptible noise and making unlearnable examples before release.
Ideally, the obtained unlearnability prevents algorithms from training usable models.
When supervised learning algorithms have failed, a malicious data collector possibly resorts to contrastive learning algorithms to bypass the protection.
Attacks need both supervised unlearnability and contrastive unlearnability.
Through evaluation, we have found that most of the existing availability attacks are unable to achieve contrastive unlearnability, which poses risks to data protection.
Furthermore, we find that employing stronger data augmentations in supervised poisoning generation can create contrastive shortcuts and mitigate this risk.
Based on this insight, we propose AUE and AAP attacks which prominently boost the worst-case unlearnability across multiple supervised and contrastive algorithms.

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: 3512