Adversarial Attack Driven Data Augmentation for Time Series Classification

Peiyu Li, Pouya Hosseinzadeh, Omar Bahri, Soukaina Filali Boubrahimi, Shah Muhammad Hamdi

Abstract: In recent years, there has been a proliferation of Time Series Classification (TSC) models that have been proposed and implemented in various real-life applications. The increase in time series availability in some scientific domains has primarily contributed to the enrichment of TSC algorithms. However, the high performance of TSC models heavily relies on a sufficiently labeled training dataset. In reality, data availability in some domains remains limited, negatively affecting the performance of TSC models. To address this challenge, we propose AGSM and AIM, aiming to enlarge the size of the training data to improve the performance of TSC. In the computer vision domain, the use of adversarial attack mechanisms to augment medical images for segmentation is very useful in improving segmentation accuracy. Following this line of work, in this paper, we propose leveraging adversarial attack mechanisms to add positive noise to generate new synthetic samples for the time series classification task. Our experimental results on 25 real-world time series data show positive impacts on improving classification performance for most datasets. Our proposed method also shows competitive performances when compared with the other seven state-of-the-art baselines.