- Conference Article
- 10.1109/nnice58320.2023.10105771
Dynamic Decision Boundaries for Improved Robustness against Adversarial Attacks
- Feb 24, 2023
- Zeyu Long
Deep neural networks is vulnerable to adversarial attacks, where an attacker crafts input data to mislead the model. Adversarial training is a common approach to improve the robustness of deep neural networks against adversarial attacks, but existing methods mainly focus on maximizing the inter-class distance and minimizing the intra-class distance of adversarial examples at the feature level, as well as keeping them as far as possible from the decision boundary. However, these methods often rely on the assumption that the model has a determined decision boundary. In this work, we propose a novel method called Uncertainty Defender (UD) and based on uncertainty estimation to improve the robustness of deep neural networks against adversarial attacks. UD models the weight vector of the classifier as a Gaussian distribution and samples from it multiple times to obtain a dynamic decision boundary. Our experimental results on CIFAR10, MNIST and ImageNet show that our method performs well in both white-box and black-box attack scenarios.
Read more