Concealed Adversarial Attacks on Neural Networks for Sequential Data

The emergence of deep learning has enabled the widespread application of neural networks in time series analysis across domains such as finance and medicine. Despite their effectiveness, these models remain vulnerable to adversarial attacks: small targeted perturbations in input data can significantly alter a classifier’s outputs. However, even formally weak attacks in the time series domain become easily detected by the human eye or a simple detector model. In this work, a concealed adversarial attack is introduced for various time series models. The approach generates realistic perturbations that are difficult for both humans and model-based discriminators to detect by jointly maximizing classifier loss and discriminator loss. A specialized training strategy is employed to enhance the discriminator’s ability to generalize across different attack strengths. Evaluation on eight benchmark datasets from the University of California Riverside (UCR) and University of East Anglia (UEA) archives, using recurrent, convolutional, state-space, and transformer-based architectures, demonstrates that the proposed method achieves a favorable trade-off between concealability and attack effectiveness. The findings highlight the growing challenge of designing robust time series models, emphasizing the need for improved defenses against stealthy and effective attacks.