Local Linearity and Double Descent in Catastrophic Overfitting
arXiv:2111.10754 · doi:10.1145/3493700.3493742
Abstract
Catastrophic overfitting is a phenomenon observed during Adversarial Training (AT) with the Fast Gradient Sign Method (FGSM) where the test robustness steeply declines over just one epoch in the training stage. Prior work has attributed this loss in robustness to a sharp decrease in of the neural network with respect to the input space, and has demonstrated that introducing a local linearity measure as a regularization term prevents catastrophic overfitting. Using a simple neural network architecture, we experimentally demonstrate that maintaining high local linearity might be to prevent catastrophic overfitting but is not Further, inspired by Parseval networks, we introduce a regularization term to AT with FGSM to make the weight matrices of the network orthogonal and study the connection between orthogonality of the network weights and local linearity. Lastly, we identify the phenomenon during the adversarial training process.
arXiv commentTo appear in 9th ACM IKDD CODS and 27th COMAD, 2022
References in corpus (6)
- Explaining and Harnessing Adversarial Examples
- Intriguing properties of neural networks
- Towards Deep Learning Models Resistant to Adversarial Attacks
- Parseval Networks: Improving Robustness to Adversarial Examples
- Understanding and Improving Fast Adversarial Training
- More Data Can Hurt for Linear Regression: Sample-wise Double Descent