Local Differential Privacy In Smart Manufacturing: Application Scenario, Mechanisms and Tools

To utilize the potential of machine and deep learning, enormous amounts of data are required. A common and beneficial approach is to share datasets between the parties involved for training purposes or even to release datasets to the public. However, several incidents have shown that despite anonymizing the data, attackers are still capable of identifying individuals in the data and extracting their sensitive information. The methods of differential privacy address this problem by adding a statistical noise to data points in the shared dataset. Since manufacturing data not only contains information about individual persons but also about the companies, their process knowledge, products, and orders add more complexity to the application of differential privacy compared to other domains. In this paper, we highlight why conventional methods of anonymization are not sufficient to guarantee data protection and thus present the necessity of using differential privacy. To illustrate its usefulness for manufacturing we present a specifc application scenario and examine potential threats when sharing manufacturing data. We identify mechanisms to perturbate data and map these to variable types in the manufacturing context. To guide practical application and research we finally outline existing differntial privacy libraries, and highlight current limitations.