LowFreqAttack: An Frequency Attack Method in Time Series Prediction

Abstract

Time series prediction has become an important research direction in data mining because of the time-varying pattern of data in various fields. However, time series prediction suffers from the problem of vulnerability to adversarial example attack, which leads to models making wrong decisions in critical application scenarios and causing great losses to people’s lives and properties. In addition, there is relatively little attacks research on time series prediction, and the existing attack methods simply migrate classical-attack methods in the image to time series prediction. On the one hand, it not only without fully considering the characteristics of temporal data but also without comparing and analyzing the effects of those classical-attack methods on time series prediction models. On the other hand, there is no comparative analysis of the effectiveness of these classical attack methods in different time series prediction methods. To address the above problems, this paper firstly compares the effectiveness of the attack methods on some time series prediction models and analyzes the inner mechanism of these time series prediction models. In addition, this paper finds that the defense ability of those models is related to their ability to portray the overall trend of time series data. Therefore, this paper further propose the new attack method, LowFreqAttack. The experimental results show that LowFreqAttack can attack the three existing time series prediction models better than the existing attach methods.