A Novel Adversarial Attack Method for Time-Series Regression Models in IIoT-Based Digital Twins

The integration of digital twin (DT) technology into the 6G-enabled Internet of Everything (IoE) has revolutionized real-time monitoring and maintenance in the Industrial Internet of Things (IIoT). However, DT models, particularly time-series regression models, are increasingly vulnerable to adversarial attacks that compromise their stability and reliability. This study proposes a reinforcement learning (RL)-based adversarial attack framework for time-series regression models, enabling the generation of highly targeted and effective adversarial examples. The method optimizes a perturbation generation strategy through RL, leveraging the temporal dynamics of time-series data to maximize its cumulative impact on the target model’s outputs under predefined perturbation constraints. Experiments on NASA’s N-CMAPSS dataset validate the method on DNN and KAN twin models using PPO and SAC algorithms, demonstrating superior attack effectiveness and stealth over FGSM, PGD, and CW, with attack intensity (AtI) scaling with perturbation magnitude. The method achieves higher computational efficiency by requiring only forward computation. Unlike gradient-based methods (e.g., APGD), the proposed approach remains effective against TRADES-trained models, showing notable adaptability. However, this advantage diminishes under hybrid adversarial training. This study exposes security risks in DT models under adversarial attacks and underscores the urgent need for advanced defense mechanisms to safeguard IoE systems.