Information Fingerprint for Secure Industrial Big Data Analytics

Data-driven models have been widely used in industry helping big data analytics due to its convenience and flexibility. However, adversarial attack has the ability to add small perturbation to actual samples, which makes the data-driven models make incorrect predictions. With the high integration of industrial control systems and information technology, the reliability and safety of data-driven models in industry have been seriously threatened. In the paper, fault diagnosis models and quality variable prediction models are verified to be vulnerable facing adversarial attack. To this end, the concept of information fingerprint containing identity information is introduced to distinguish actual samples from adversarial samples with small perturbation. With fault diagnosis models and quality variable prediction models as the background, information fingerprint exaction networks (IfeNet) is developed to extract the information fingerprint for further analysis. IfeNet utilizes supervised contrastive pre-training and unsupervised training to realize parameter learning, with the structure of Siamese deep neural network and autoencoder. Finally, the effectiveness and feasibility of the proposed approach for fault diagnosis models and quality variable prediction models are verified in two industrial benchmark cases.