Information-theoretically Safe Bias Classifier Against Adversarial Attacks
Keywords: Information-theoretically, theory, adversarial
Abstract: Deep learning has become the cornerstone of recent advances in artificial intelligence. However, the presence of adversarial samples makes deep learning susceptible in applications where safety is critical. Moreover, adversarial examples have been shown, to some degree, to be unavoidable. To address this issue, we propose the bias classifier. This approach employs the bias component of a neural network, using ReLU as its activation function, as a classifier. The bias classifier has been shown to universally approximate any classification problem with a high degree of probability. Moreover, it can be made information-theoretically safe against the original model gradient-based attack in the sense that any such attack produces a completely random attacking direction for any given input. Thus, the bias classifier provably achieves the maximum possible robust accuracy under specified attacks. Experiments are used to validate our theoretical results and to show that the bias classifier is accurate and robust for simple models.
Primary Area: learning theory
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/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 2380
Loading