| Author | Name Surname |
| Consultant | Name Surname, PhD/DSc |
| Advisor | Name Surname, PhD/DSc |
Deep learning methods are widely used for time series analysis in domains such as healthcare, finance, energy systems, and environmental monitoring. However, deep neural networks remain vulnerable to adversarial attacks, where small input perturbations can cause significant degradation in predictive performance. Commonly used gradient-based adversarial attacks (e.g., iterative first-order methods) are computationally burdensome. They require repeated backpropagation through the victim model to compute input gradients, and many iterative refinement steps to obtain high-quality perturbations under the chosen constraint set.
In this work, we propose model-based adversarial attacks in which perturbations are generated by a neural network. During training, the generative model learns to produce perturbations that maximize the loss of a frozen target model. We introduce both single-step and iterative generative attack schemes and evaluate them on multiple time series datasets and model architectures. Experimental results demonstrate that the proposed approach achieves performance comparable to classical gradient-based attacks while offering substantial computational advantages.
If you find our work helpful, please cite us.
@article{citekey,
title={Title},
author={Name Surname, Name Surname (consultant), Name Surname (advisor)},
year={2025}
}Our project is MIT licensed. See LICENSE for details.