基于归一化流的自适应Epsilon对抗训练用于鲁棒引力波参数估计
机器学习
2024-12-18 v2
摘要
基于归一化流模型的对抗训练是一个新兴的研究领域,旨在通过对抗样本提高模型鲁棒性。在本研究中,我们将对抗训练应用于归一化流模型,用于引力波参数估计。我们提出了一种用于快速梯度符号法(FGSM)对抗训练的自适应Epsilon方法,该方法通过对数缩放基于梯度幅值动态调整扰动强度。我们的混合架构结合了ResNet和逆自回归流,在FGSM攻击下将负对数似然(NLL)损失相比基线模型降低了47%,同时在干净数据上保持了4.2的NLL(仅比基线高5%)。在0.01至0.1的扰动强度范围内,我们的模型实现了平均NLL为5.8,优于固定Epsilon(NLL: 6.7)和渐进Epsilon(NLL: 7.2)方法。在更强的投影梯度下降攻击下,扰动强度为0.05时,我们的模型保持了6.4的NLL,展现出优越的鲁棒性,同时避免了灾难性过拟合。
引用
@article{arxiv.2412.07559,
title = {Adaptive Epsilon Adversarial Training for Robust Gravitational Wave Parameter Estimation Using Normalizing Flows},
author = {Yiqian Yang and Xihua Zhu and Fan Zhang},
journal= {arXiv preprint arXiv:2412.07559},
year = {2024}
}
备注
Due to new experimental results to add to the paper, this version no longer accurately reflects the current state of our research. Therefore, we are withdrawing the paper while further experiments are conducted. We will submit a new version in the future. We apologize for any inconvenience this may cause