Robust Multivariate Time-Series Forecasting: Adversarial Attacks and Defense Mechanisms

Explore the challenges and solutions in robust multivariate time-series forecasting through this comprehensive lecture on adversarial attacks and defense mechanisms. Delve into the research of Dr. Hoang, an Assistant Professor at Washington State University, as he presents findings from his work published at ICLR-23. Discover a newly identified attack pattern that impacts target time series forecasting through strategic, sparse modifications to past observations of other time series. Learn about two innovative defense strategies: an extension of randomized smoothing techniques to multivariate forecasting scenarios, and an adversarial training algorithm that simultaneously creates adversarial examples and optimizes the forecasting model for improved robustness. Gain insights from extensive experiments on real-world datasets that demonstrate the power of these attack schemes and the effectiveness of the proposed defense algorithms compared to baseline mechanisms.