基于生成对抗网络的对抗防御系统

针对不断更新的对抗攻击,提出一个基于生成对抗网络的防御系统.系统利用生成对抗网络不断生成新的对抗样本,反复训练模型以增强其鲁棒性.具体过程为将预先训练的卷积神经网络和外部GAN(conditional GAN:Pix2Pix)相结合,自动流水线式地推断对抗样本和干净样本之间的转换关系,并合成新的对抗样本.根据分辨得到的...

Adversarial Defense System Based on Generative Adversarial Network

To address the evolving adversarial attack, this paper proposes a defense system based on generative adversarial network. To enhance its robustness, the system uses the generative adversarial network to generate new adversarial samples to train the model repeatedly. The specific process is to combine the pre-trained convolutional neural network with the external Gan ( conditional Gan:pix2pix) to automatically infer the conversion relationship between adversarial samples and clean data,and synthesize new adversarial samples. The generator and discriminator in the generative adversarial network are adjusted continuously according to the feedback result of discrimination to enhance its performance, while the newly synthesized adversarial samples are used to strengthen the defense model in the iterative pipeline. Finally, the effectiveness of the system is proved by experiments.