Adversarial Attacks on Deep Neural Networks for Time Series Prediction

Abstract

Time series data is widespread in real-world scenarios. To recover and infer missing information in practical domains, such as stock price monitoring, electricity load forecasting, traffic flows analysis, climate trend prediction, etc., the problem of time series prediction has been widely studied as a classical research topic in data mining. Over the past decade, deep learning architectures are introduced as a vital part of the next generation of time series prediction models. However, recent studies showed that deep learning models are vulnerable to adversarial attacks. In this paper, we study the adversarial attacks on the time series prediction models prospectively. We propose an attack strategy to generate adversarial samples by adding imperceptible perturbed data to the original time series with the goal of reducing the accuracy of time series prediction models. Specifically, the perturbation-based adversarial example generation algorithm is proposed using gradient information of time series prediction model. Moreover, adversarial examples should be imperceptible to humans. To address the challenge, we craft adversarial samples based on importance measuring to perturb the original data locally. We evaluate our attacks on state-of-the-art time series prediction models using three time series datasets. Our results demonstrate that our attacks can effectively evade the time series prediction models, and the adversarial attacks mechanisms can be used as robustness metric for constructing robust time series prediction models.