A Query-less Adversarial Attack Method Against Time Series Forecasting

Adversarial attacks in Time Series Forecasting (TSF) have become a topic of growing interest in recent years. However, most previously proposed black-box attack methods against TSF required a vast number of queries to the target model to ensure effective attack performance. In our approach, we aim to learn from adversarial examples and predict the sensitive locations within the data. Specifically, we collect adversarial samples using intelligent optimization methods, such as Particle Swarm Optimization (PSO), to identify locations that can be perturbed in the adversarial examples. Furthermore, we train a model to establish a connection between the original data and its corresponding attack locations. Finally, this trained model is utilized to predict sensitive locations and generate new adversarial examples. The experimental results demonstrate that our method achieves comparable performance to original attack methods while significantly reducing the number of queries to the target model.