Protecting the interests of owners of intelligent fault diagnosis models: A style relationship-preserving privacy protection method

Abstract

The training process of intelligent fault diagnosis models requires data collection and computational cost, which is laborious and time-consuming. Due to the similarity of the fault categories to be classified between different models, an unauthorized user may easily transfer a diagnosis model to his own domain, infringing the interests of the model owner. Therefore, it is important to protect the intellectual property of fault diagnosis models. The existing privacy protection methods use a metric to increase the feature divergence or construct un-transferable isolation domains, but the generalization boundaries of the deep learning models cannot be minimized enough, leading to certain availability in unauthorized domains. To tackle this issue, a style relationship preserving based model property protection method is proposed for fault diagnosis in this paper. The proposed method utilizes style transfer technique to simulate the extreme situation, which further compacts the generalization areas and limits the application of the model to unauthorized domains. Comprehensive experiments are established on a public fault diagnosis dataset and a practical rolling bearing fault diagnosis test platform, demonstrating the effectiveness and superiority of the proposed method in model privacy protection.