Robustness of Unsupervised Representation Learning without Labels
This work addresses robustness evaluation for unsupervised learning, which is incremental by extending existing adversarial methods to a label-free setting.
The paper tackles the problem of evaluating robustness in unsupervised representation learning without relying on downstream tasks, proposing label-free measures and adversarial training extensions that improve certified accuracy and reduce attack success rates, such as achieving 3 times higher certified accuracy for MOCOv2.
Unsupervised representation learning leverages large unlabeled datasets and is competitive with supervised learning. But non-robust encoders may affect downstream task robustness. Recently, robust representation encoders have become of interest. Still, all prior work evaluates robustness using a downstream classification task. Instead, we propose a family of unsupervised robustness measures, which are model- and task-agnostic and label-free. We benchmark state-of-the-art representation encoders and show that none dominates the rest. We offer unsupervised extensions to the FGSM and PGD attacks. When used in adversarial training, they improve most unsupervised robustness measures, including certified robustness. We validate our results against a linear probe and show that, for MOCOv2, adversarial training results in 3 times higher certified accuracy, a 2-fold decrease in impersonation attack success rate and considerable improvements in certified robustness.