GENADA: Efficient Generative Adversarial Attacks on Time Series Models

Deep learning models are increasingly used for time series analysis across various critical sectors like healthcare and finance. However, these models are susceptible to adversarial attacks, where minor input changes can significantly degrade their performance. Traditional gradient-based attack methods are computationally intensive, requiring multiple backpropagation steps to generate perturbations. A new framework, GENADA (GENerative ADversarial Attack), addresses this by learning a generative model to produce these deceptive perturbations directly. This allows for the creation of adversarial examples in a single forward pass, significantly reducing the computational burden during inference. The framework includes both single-step and iterative generative attack schemes. Validation experiments on various neural models and datasets in low-dimensional time-series settings demonstrated that GENADA achieves attack quality comparable to strong baseline methods. Crucially, it does so while requiring substantially less time to generate the adversarial perturbations during the inference phase. This efficiency makes it a promising approach for evaluating and improving the robustness of time series models.