Secure Machine Learning in the Cloud

Purpose and goal:

The aim of this project was to investigate implementation of secure machine learning algorithms using homomorphic encryption as cloud services, including analysis of practical and policy challenges. We have proposed an architecture and implemented a PoC, showing that homomorphic encryption is a promising technology. However, there are limitations in how the algorithms can be implemented. Also, rules and laws for sharing sensitive data (e.g., banking, patient, or governmental data) are unclear.

Expected results and effects:

We have shown the potential of secure machine learning using homomorphic encryption in cloud services. We have developed a PoC in which we have investigated a potential service architecture including its opportunities and limitations. The results have led to further developments around homomorphic encryption (e.g., regarding scalability and complexity) as well as its application in other areas, based on signs of interest from other industrial sectors during the project.

Approach and implementation:

The project was implemented in work packages for research, platform development, software development, use-case, and collaboration. The work was synchronized in regular meetings and several workshops. An iterative approach was employed where basic functionality was developed first, followed by more complex functionality. This resulted in tight collaboration between the partners. A challenge arose due to a staffing issue after half of the project, which required prolongation to finish it.