An Adversarial Attack Method for Multivariate Time Series Classification Based on AdvGAN

Considering the complexity of time series data and real-world applications, multivariate time series classification models are vulnerable to adversarial attacks. Although existing white-box attack strategies have made progress in generating adversarial samples, they rely on access to the target model’s parameters, training data, and gradients. Therefore, we apply AdvGAN framework for multivariate time series classification. AdvGAN is designed as a framework based on Generative Adversarial Networks (GANs), encompassing a generator, discriminator. The generator creates multivariate perturbations, and the perturbations combine with original data to form adversarial samples. The discriminator assesses the authenticity of these samples. These samples are then used to evaluate the security of the target model. We conducts experiments across three University of East Anglia (UEA) and University of California Riverside (UCR) datasets, employing the Multivariate Long Short Term Memory Fully Convolutional Network (MLSTM_FCN) as the target model for adversarial attack testing. The results indicate that our designed attack method effectively enhances the success rate of adversarial attacks while maintaining a similar level of Mean Squared Error (MSE) between the generated adversarial samples and the original samples.