Transfer adversarial attacks across industrial intelligent systems

As indispensable parts of industrial production control, data-driven industrial intelligent systems (IIS) achieve efficient executions of significant tasks such as fault classification (FC), fault detection (FD), and soft sensing (SS). Recently, machine learning models have been proven vulnerable to adversarial attacks, where the transfer-based attacks provide highly feasible attacks on systems in real-world black-box scenarios. In this paper, to study the practical security risks of IIS, we investigate transferable adversarial attacks from: (1) showing the existence of transferable adversarial examples across different industrial tasks; (2) exploring factors (e.g., data feature, model structure, and attack method) affecting transferability under multi-scenarios; (3) proposing a new method to enhance the transferability; (4) providing guidelines on practical system deployments to defend against transferable adversarial threats. The attacks demonstrate generality on two types of datasets, Tennessee Eastman industrial process (TEP) and WM-811K wafer map dataset, and the experiment results show that: (1) transfer is asymmetric and complex models are relatively stable with low sample transferability; (2) iterative and single-step methods have opposite performance characteristics under the intra- and cross-task transfer; (3) overfitting of optimization methods leads to weak transferability; (4) smoothing gradients and widening intermediate layer perturbations are effective directions for improving transferability.