A corresponding project was launched under the direction of Fraunhofer Austria. The consortium is investigating the practicality of machine learning on encrypted data. […]
There is enormous potential in the data that companies collect and store in ever greater quantities, which can be unlocked with machine learning methods. On the one hand, machine learning with large amounts of data works particularly well and, on the other hand, it is not worthwhile for every company to build up know-how and infrastructure for the application of corresponding methods, so that companies would benefit from cooperation. When it comes to sensitive data, however, this is only an option for many companies if appropriate security measures are taken. Methods that allow secure processing of data have been researched for years. Previous results show that a certain type of encryption is, at least in theory – to enable secure machine Learning on sensitive data. A consortium of several companies and research institutions in the FFG SMiLe project, headed by Fraunhofer Austria, now wants to examine whether and how exactly this method can also be used meaningfully in practice. The kick-off for the project took place in April.
Daniel Bachlechner, Head of the Advanced Data Analytics research group at Fraunhofer Austria and coordinator of the SMiLe research project, describes the initial situation as follows: “Especially for applications that are based on machine learning and thus particularly benefit from large amounts of data, a combination of data across organizational units and company boundaries would be important. Nevertheless, data is hardly shared today and – if at all-used primarily where it was collected. In order to be able to actually use the full potential of sensitive data, data must at the same time be protected from unauthorized access and used as fully as possible for calculations.”According to the researchers, a promising approach to achieving this goal is so-called homomorphic encryption.
“The method is most comparable to an opaque glove box. This means that someone can put objects in and lock the box, and someone else – by putting their hands in their gloves-can work with the objects in the box without seeing them,“ explains Michael Rader, research associate at Fraunhofer Austria. It has already been shown that calculations can actually be carried out on such encrypted data, but in practice there is a lack of know-how and suitable software. The research team now wants to change this: SMiLe addresses both aspects and thus wants to create an essential prerequisite for the practical use of machine learning on encrypted data. Michael Rader explains: “In the last ten years, new techniques have been added that have made the method more practical and a lot has happened in terms of computing power and hardware support. We will now consider whether it is already possible to solve those problems with it that one faces in practice.“
The requirements for the processing and protection of data are diverse and vary from company to company. It is therefore important for the researchers to investigate various potential applications in the course of the project. With Fill, a company is on board that wants to make machine data usable for predictive maintenance on the basis of a technical platform provided by Tributech, without disclosing sensitive data from which the process know – how of the machine operators could be reconstructed. CORE smartwork, on the other hand, focuses on the use of sensitive employee data. Using machine learning, the company wants to support its customers in assembling teams that not only ensure efficient task fulfillment, but also allow work-integrated training of employees. Patrick Lamplmair, CTO at Tributech, sees homomorphic encryption as an enabler for data services in trusted ecosystems that can also be used to leverage highly sensitive data along the value chain. “Tributech’s goal is to offer the developed solution as an extension of the DataSpace kit in the future and thus make the technology easy to use for end users,” explains Patrick Lamplmair.
But not only the applications are diverse, but also the expertise required to transfer the theoretical preparatory work into practice: While Fraunhofer Austria brings experience with machine learning to the project, the Management Center Innsbruck provides the necessary knowledge about cryptographic methods. The Software Competence Center Hagenberg and the VRVis Center for Virtual Reality and Visualization also make important contributions to the project with their expertise in the areas of explainable artificial intelligence and data visualization. Pascal Schöttle, associate Professor at the Management Center Innsbruck assumes that the importance of security in connection with machine Learning will continue to increase. In addition to insights into the practicality of homomorphic encryption, Pascal Schöttle expects new theoretical insights from the project, but is also aware of the challenges that need to be overcome: “One of the biggest challenges in the course of SMiLe will probably be to efficiently combine an extremely compute-intensive application, such as machine learning, with an equally compute-intensive form of encryption, as is the case with homomorphic encryption.“
About the SMiLe research project
The project “Secure Machine Learning Applications with Homomorphically Encrypted Data” (SMiLe) is funded under the project number 886524 by the Federal Ministry for Climate Protection, the Environment, Energy, Mobility, Innovation and Technology (BMK) as part of the 8th call for proposals of the “ICT of the Future” program. Fraunhofer Austria is implementing the project together with the Management Center Innsbruck, the Software Competence Center Hagenberg and the VRVis Center for Virtual Reality and Visualization as well as the companies Fill, CORE smartwork and Tributech. The project has been running since April 2021 and will end in September 2023 after a duration of 30 months.