Directed Infusion of Data (DIOD) for Secure Data Transfer
Primary Investigator:
Hany Abdel-Khalik
Tyler Lewis, Arvind Sundaram
Abstract
The emergence of AI/ML tools heavily incentivizes collaboration among industrial partners and research institutions to ensure that the vast quantities of data they possess are efficiently leveraged. Large dynamic systems such as power plants, grid applications, smart factories, etc., require extensive engineering analytics such as condition monitoring, time series prediction, control implementation, etc. However, the possibility of data leakage due to a malicious third-party or an untrustworthy collaborator leads to major security concerns that may prevent collaboration. While most collaborators are assumedly trustworthy, it is difficult to absolutely ensure that data cannot be misused; even when the analyst is trustworthy, sensitive data passed to them is vulnerable to their network’s security. While various researchers have experimented with techniques to bypass this assumption, the current state-of-the-art methods inherently limit the results of collaboration, inducing a tradeoff between data privacy and fidelity of results. DIOD allows efficient data masking that cannot be reverse engineered by any third-party, regardless of their knowledge of the system. In addition, the data’s inferential properties are preserved by DIOD, permitting complex analyses, (e.g., classification, regression, etc.), to be performed on the masked data directly. DIOD has been studied across a variety of circumstances, including condition monitoring tasks, simple binary classification problems, and linear regression.