基于对抗训练增强模型鲁棒性的新方法

Adversarial Training Method Based on Random Perturbation

将对抗样本引入训练过程可以提高深度学习模型的鲁棒性,而且能为模型提供可解释的梯度,但这一防御策略往往需要较多的计算资源和时间成本。为提升训练模型效率和鲁棒性,同时降低训练成本,提出一种基于随机扰动的对抗训练方法:首先利用基于FGSM(Fast Gradient Sign Method)的随机扰动方法生成对抗样本;其次,所提出算法的优越性;最后,利用周期性学习率和Adam(Adaptive Moment Estimation)相结合方法更新学习率。实验结果表明,通过引入周期性学习率机制,整个对抗训练过程的稳定性和拟合效果有了显著提升,所提出的训练方法能降低训练成本和提高模型的性能。

Introducing adversarial examples into the training process can improve the robustness of deep learning model and provide interpretable gradients for the model,but this defense strategy often requires more computational resources and time cost.In order to reduce the training cost while improving the efficiency and robustness of the trained model,proposes a random perturbation-based adversarial training approach.First,the random perturbation method based on FGSM(Fast Gradient Sign Method)is used to generate adversarial examples;then,the superiority of the proposed algorithm is analyzed,finally,the learning rate is updated using a combination of periodic learning rate and Adam(Adaptive Moment Estimation)optimizer.The experimental results show that the stability and the fitting effect of the whole adversarial training process are significantly improved by introducing the periodic learning rate mechanism,and the training approach proposed can reduce the training cost and improve the performance of the model.