CASTLEGUARD: Anonymised Data Streams with Guaranteed Differential Privacy
Alistair Robinson, Frederick Brown, Nathan Hall, Alex Jackson, Graham Kemp, Matthew Leeke · 2020
Data streams are commonly used by data controllers to outsource the processing of real-time data to third-party data processors. Data protection legislation and best practice in data management support the view that data controllers are responsible for providing a guarantee of privacy for user data contained within published data streams. Continuously Anonymising STreaming data via adaptive cLustEring (CASTLE) is an established method for anonymising data streams with a guarantee of k-anonymity. However, k-anonymity has been shown to be a weak privacy guarantee that has vulnerabilities in practical applications. In this paper we propose Continuously Anonymising STreaming data via adaptive cLustEring with GUAR-anteed Differential privacy (CASTLEGUARD), a data stream anonymisation algorithm that provides a reliable guarantee of k-anonymity, l-diversity and differential privacy to data subjects. We analyse CASTLEGUARD to show that, through safe k-anonymisation and β-sampling, the proposed approach satisfies differentially private k-anonymity. Further, we demonstrate the efficacy of the approach in the context of machine learning, presenting experimental analysis to demonstrate that it can be used to protect the individual privacy of users whilst maintaining the utility of a data stream.