今天读了一下 Towards Evaluating the Robustness of Neural Networks,这是关于对抗样本(Adversarial Example)的paper,主要贡献是提出了Carlini & Wagner Attack神经网络有目标攻击算法,打破了最近提出的神经网络防御性蒸馏(Defensive Distillation),证明防御性蒸馏不会显著提高神经网络的鲁棒性,论文的信息量还是比较大的。
可能需要阅读的论文以获取前置知识:
关于蒸馏网络:
[1] Do Deep Nets Really Need to be Deep?
[2] Distilling the Knowledge in a Neural Network
[3] Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks
关于对抗样本:
[4] Intriguing properties of neural networks(L-BFGS Attack)
[5] Exploring the Space of Adversarial Images(FGSM)
[6] Towards Deep Learning Models Resistant to Adversarial Attacks(PGD, I-FGSM)
[7] The Limitations of Deep Learning in Adversarial Settings (JSMA)
[8] DeepFool: a simple and accurate method to fool deep neural networks(DeepFool)
关于论文源码: