Tag
该论文从理论上解释了对抗训练的神经网络其梯度显著性图呈现稀疏性的现象：对于两层ReLU网络，ℓ∞对抗攻击导致梯度范数各向异性，从而偏好轴对齐/稀疏的梯度。研究在经验风险最小化框架下证明了最小化收敛到具有最小梯度和Barron范数的贝叶斯分类器。
Introduces a weight perturbation-based feature attribution method (XWP and XWPc) for fully connected neural networks, achieving competitive performance on standard baseline metrics.