Summary
International Workshop on Smart Info-Media Systems in Asia
2023
Session Number:RS3
Session:
Number:RS3-1
Evaluating Effectiveness of Adversarial Examples for Multivariate Time-series Data with Externality
Masatomo Yoshida, Masahiro Okuda,
pp.66-70
Publication Date:2023/8/31
Online ISSN:2188-5079
PDF download
Summary:
Re-training with adversarial examples is often used to improve the robustness of machine learning and to address vulnerabilities, especially in image recognition. On the other hand, research on the adversarial perturbations for time-series data is not as advanced as for images. This is due to the difficulty of acquiring large amounts of data and the high cost of artificially generating data, despite its applicability to familiar and important areas, such as energy supply and demand, weather, and economic indicators. Data augmentation using adversarial perturbation can effectively address these issues for time-series data. In this paper, we examine the effectiveness of adversarial perturbation on multivariate time series data with externalities. We first define multivariate time-series data with externality, and then we use actual Japanese stock market data to verify the accuracy of forecasting using adversarial perturbation. Next, we focus on the possibility of contradictory items in applying adversarial perturbations and propose a constrained perturbation that accounts for such contradictions, and confirm its effectiveness.