Dynamic Time Warping Based Adversarial Framework for Time-Series Domain.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep neural networks (DNNs) show rapid progress in adversarial robustness, but lack principled methods for time-series data.
• Time-series data is crucial in mobile health, finance, and smart grids, necessitating robust DNNs.

Purpose of the Study:

• To propose a novel framework, Dynamic Time Warping for Adversarial Robustness (DTW-AR), for enhancing DNN robustness in the time-series domain.
• To demonstrate the superiority of Dynamic Time Warping (DTW) over Euclidean distance for time-series adversarial attacks.

Main Methods:

• Developed a novel framework, DTW-AR, leveraging the dynamic time warping measure for time-series adversarial robustness.
• Designed a principled algorithm, justified by theoretical analysis, for efficiently generating diverse adversarial examples using random alignment paths.
• Evaluated the framework on diverse real-world time-series benchmarks.

Main Results:

• DTW-AR effectively fools DNNs on time-series data, outperforming standard Euclidean distance-based methods.
• Adversarial training with DTW-AR significantly improves the robustness of DNNs against time-series attacks.
• Theoretical and empirical evidence supports the effectiveness of DTW for time-series adversarial robustness.

Conclusions:

• DTW-AR provides a principled and effective approach to adversarial robustness for deep neural networks in the time-series domain.
• The proposed framework enhances the security and reliability of DNNs in critical time-series applications like mobile health and finance.