Interpreting Adversarially Trained Convolutional Neural Networks

个人总结： 其实说白了就是： AT-CNNs are more biased towards global structures, such as shapes and edges。        AT-CNN对图像的整体结构，比如形状，边缘更加敏感，可以更好的捕捉到这些信息。 AT-CNNs are less sensitive to the texture distortion and foc

>>阅读原文<<