FABLE:面向天气预测模型的局部、定向对抗攻击
机器学习
2026-04-07 v2 密码学与安全
摘要
深度学习基于天气预测(DLWF)模型最近在性能上显著优于金标准物理仿真工具,但这些模型可能对对抗攻击存在脆弱性,这引发了可信度问题。本文研究了将现有对抗攻击方法应用于DLWF模型的可行性与挑战,并提出一种新框架称为FABLE(Forecast Alteration By Localized targeted advErsarial attack),以解决上述问题。FABLE对数据进行3D离散小波分解,以分离其空间和时间分量。通过调节不同分量中对抗扰动的幅度,FABLE生成的对抗输入保持与原始输入高度一致,同时引导DLWF模型生成定向的预测结果。在真实世界气象数据集上进行的实验结果表明,FABLE在各种指标上均优于基线方法。
引用
@article{arxiv.2505.12167,
title = {FABLE: A Localized, Targeted Adversarial Attack on Weather Forecasting Models},
author = {Yue Deng and Asadullah Hill Galib and Xin Lan and Jack Gunn and Pang-Ning Tan and Lifeng Luo},
journal= {arXiv preprint arXiv:2505.12167},
year = {2026}
}
备注
Version 2 incorporates revisions based on feedback from NeurIPS 2025 reviewers (final score: borderline). We improved clarity in previously complex sections to enhance accessibility for non-expert readers and expanded the experimental evaluation to provide more comprehensive and diverse results