Research on Adversarial Disturbance Based on Meteorological Time Series Data

Abstract

When the deep learning model is used to predict time series data, it is easy to be adversarially attacked. The time series data is sensitive to the abnormal disturbance and has strict requirements on the disturbance amount. To solve these problems, we propose to generate adversarial time series by adding disturbance terms to the original time series, and design an adversarial attack algorithm based on the importance measure (AAIM in short), which slightly perturbs the original data to improve the performance of the time series prediction model. The meteorological data of Guangzhou from 1980 to 2020 are selected as the studied time series data, which verifies that the proposed adversarial attack method not only effectively fools the target time series prediction model LSTNet, but also attacks the most advanced models based on CNN and RNN. It can be seen from the results of AAIM that the adversarial disturbance based on the first 5% importance measurement has almost the same effect as the 100% disturbance of the original time series. For the proposed AAIM, even if the disturbance is tiny, it can impact the prediction performance of the deep learning model significantly. Experimental results show that the proposed method achieves good transferability.