CAT: Collaborative Adversarial Training
Keywords: adversarial training, adversarial robustness
Abstract: Adversarial training has proven to be effective in enhancing the robustness of neural networks. However, previous methods typically focus on a single adversarial training strategy and do not consider the characteristics of models trained using different strategies. Upon revisiting these methods, we have observed that different adversarial training methods exhibit distinct levels of robustness for different sample instances. For instance, a model trained using AT may correctly classify a sample instance that is misclassified by a model trained using TRADES, and vice versa. Motivated by this observation, we propose a Collaborative Adversarial Training (CAT) framework to enhance the robustness of neural networks. CAT utilizes different adversarial training methods to train robust models and facilitate the interaction of these models to leverage their combined knowledge during the training process.Extensive experiments conducted on various networks and datasets validate the effectiveness of our method.
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
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: 599
Loading