Knowledge-Aided Generative Adversarial Network: A Transfer Gradient-Less Adversarial Attack for Deep Learning-Based Soft Sensors

With the increasing complexity of modern industrial and process data, deep learning-based soft sensor (DLSS) models have provided more efficient sensing results than those of traditional soft sensors, but their vulnerability to adversarial attacks has hindered their deployment and safe application. Because it is difficult for the existing DLSS attack methods to accurately estimate the transfer gradient, this study proposes an adversarial attack method called knowledge-aided generative adversarial network (KAGAN) to directly generate adversarial samples by using deep neural networks, thereby avoiding the estimation and use of the transfer gradient. Firstly, the KAGAN utilize the proposed automatic peak-finding algorithm to get the auxiliary knowledge from the modeling data of the original DLSS. The auxiliary knowledge is the fake sensing results expected to be realized by the attack. In the KAGAN structure, a GAN and a proxy soft sensor model are fused, and the loss function of this structure is redesigned based on the set fake sensing results. After the KAGAN is trained and convergent, the generator part of the KAGAN is extracted and used to directly generate adversarial samples. The effectiveness of the proposed attack method is verified using the thermal deformation sensing data of an air preheater rotor. The experimental results show that the KAGAN attack is superior in terms of imperceptibility, rationality, and stability than the transfer gradient-based attack methods, thus proving that DLSS attacks can be realized without relying on the transfer gradient estimation. Moreover, KAGAN can reveal additional vulnerabilities of DLSSs in the face of adversarial attacks, thereby laying the foundation for further research on defense methods, which is crucial for improving the security of DLSSs.