The University of Amsterdam wants to make privacy-safe student and employee data available for research and analysis purposes.
Representative data on study and career paths is a key ingredient for fair policy making
This project is part of the Amsterdam Data Exchange (AMdEX) innovation fieldlab where parties work together to develop and test services to translate real world policies for sharing data into digital agreements. These can be digitally enforced, monitored and controlled: enabling parties that wish to share data to be compliant with law, regulations and consortium policies. Within this fieldlab the University of Amsterdam (UvA), Universities of The Netherlands (UNL) and BlueGen.ai explored the application of synthetic data in a collaborative research environment and the impact on digital agreements.
In the context of the UNL, universities show the outside world how they shape their social task, formulate joint ambitions regarding scientific education and research and lobby for the preconditions that are necessary to realize these ambitions. The UNL office facilitates this, among other things, by making figures and data about the university sector available and offering a platform for knowledge exchange.
One of the topics within UNL is the equal opportunities for all students and employees. The underlying data needed for analysis is privacy-sensitive and is subject to General Data Protection Regulation (GDPR). This makes it difficult to do collaborative research and combine the data of multiple universities.
Fourteen universities work through education and research with over 61 thousand employees and over 340 thousand students to build a strong knowledge society that contributes to innovation as well as solving scientific questions from society.
The University of Amsterdam leads the way
The Centre for Digital Innovation and Artificial Intelligence at UvA (LAB42) has taken the initiative to solve the above challenge, because teaching, social responsibility and innovation are considered just as important as research at the UvA. It should be a place where a diverse range of people, experiences and ideas come together. Students and employees need to feel at home and receive equal opportunities.
The main goal of this project is to apply synthetic data as a privacy-enhancing technology and enable research to improve policies for equal opportunities. The work is based on data from the Scientific Education HR System (WOPI) and several sub goals are defined:
- assess the quality of synthetic data in the context of WOPI research
- familiarize with the process and requirements of synthetic data generation
- position synthetic data in the AMdEX infrastructure and agreements
Artificial intelligence to generate generate synthetic data
With BlueGen.ai’s software, UvA is able to generate synthetic versions of data sets that are no longer traceable to confidential and personal information from individual records. The software runs on Artificial Intelligence that learns from the original data and creates synthetic data sets that look and behave the same as the underlying real data. They are suitable for experimental research and machine learning development, without the risk of leaking sensitive information. This gives researchers and policy makers at UvA and UNL the ability to use the statistics of WOPI data that were previously not accessible.
A privacy-by-design solution to minimize risks
Employee data is highly sensitive and easy to link to other personal data. In consultation with the Privacy Officer, various safety and security related measures have been taken to handle the data with the utmost care. The BlueGen.ai software is installed as a stand-alone Docker container without internet access on UvA’s secure environment of the LAB42 MultiScale Networked Systems group. Both the real and synthetic data sets are only accessible by local authorized UvA researchers and the real data is deleted from the secure environment after training the model.
Assessing the quality and privacy of synthetic data
Researchers from UvA first used the standard evaluation capabilities of the BlueGen.ai software to assess the synthetic data set. This includes various distribution, similarity, correlation, temporal and privacy metrics. The latter is based on work from WP29 and IEEE that define the following:
“an effective anonymisation solution prevents all parties from singling out an individual in a dataset, from linking two records within a dataset (or between two separate datasets) and from inferring any information in such dataset.”
After the initial quality and privacy evaluation based on BlueGen.ai software, the UvA researchers asked domain specific questions to both the real and synthetic data sets, to determine the fit-for-purposeness. The answers to these questions are confidential and can not be published, but included mean salary scales and salary growths per gender over time for various roles.
“BlueGen.ai’s synthetic data provides similar answers to statistical analyses on WOPI data”
Ana Oprescu,
Assistant Professor LAB42 – University of Amsterdam
Synthetic data in practice
The University of Amsterdam has proven that it is possible to generate a privacy-safe and representative synthetic WOPI data set with the BlueGen.ai software. The standard resemblance, utility and privacy evaluations which are supplemented by the domain specific analyses from UvA experts give enough trust to continue the exploration of synthetic data in practice. Especially for internal university research topics around students and employees, synthetic data is a viable option.
Taking it one step further is to generate synthetic WOPI data over multiple universities under the UNL umbrella and do collaborative exploratory research for policy making. Synthetic data solves a big part of the privacy challenge when comparing and combining insights from multiple universities. The conditions under which either real or synthetic data is available can be enforced, monitored and controlled by a digital notary (AMdEX), while synthetic data drastically lowers those requirements for data sharing. Besides that, the BlueGen.ai solution can run either in a compute to data (at the university) or trusted third party environment like SURF which is the IT cooperative of education and research in The Netherlands.
- You want to research the statistics in an external data set without a very specific scope
- You want to experiment with your own models and software on an external data set
- You want to enrich external research with the statistical value of your own data set, but not share the individual records due to confidentiality or privacy