More and more businesses across several industries are interested in adopting AI, and there is an increasing number of providers of these services. As a provider themselves, Midlaier AS sees a lack of clear guidelines and good methods for implementing AI in compliance with existing and upcoming laws and regulations (e.g., the EU AI Act) in applications that process personal data or other sensitive information. This is an important piece in creating trust with customers. Additionally, we recognise that while significant advances are being made in new ways to apply advanced AI, security and reliability often take a backseat to new functionality. Privacy and data governance is one of the seven pillars in the EU's definition of trustworthy AI (EU, 2019).
In a machine learning scenario, multiple parties must collaborate. One party typically owns the training data, another set of parties may own the data used for inference, and a third party might offer an AI application that performs the inference. Furthermore, it operates on an infrastructure sourced from a long supply chain with many obscure levels of abstraction (e.g., cloud services), involving many parties. This trust network makes it difficult for both the actors purchasing the services and the users, whose data is being processed, to have enough confidence in such systems to exploit their full potential. With this project, we aim to work towards solutions where there is platform-agnostic security and privacy at every stage of this data processing chain.
The project "Trustworthy and Privacy-Preserving Artificial Intelligence" aims to contribute to developing methods and technologies for ensuring privacy and data security in systems that train and employ AI technology. The project aims to help increase the trust of the business community, the public sector, and society as a whole in safe and reliable AI so that the technology's vast potential can be used for many beneficial societal purposes.
Prosjektet «Pålitelig og personvernbevarende KI» (Eng.: “Trustworthy and privacy-preserving AI”) har som hovedmål å utvikle nye metoder og teknologier som sikrer personvern og datasikkerhet i systemer som er basert på, eller benytter seg av kunstig intelligens (KI). Midlaier AS leverer i dag KI-baserte digitale assistentløsninger til både privat og offentlig sektor og er involvert i flere pågående og planlagte FoU-løp innen både justis-, helse- og forsvarssektorene. Selskapet jobber blant annet med å utvikle løsninger for assisterte og effektiviserte pasientintervju, en assistent for HR-arbeid i Forsvaret og en applikasjon for øvelsessimulering tiltenkt både offentlig og privat sektor. I alle disse applikasjonene er sikkerhet i både behandling og lagring av personopplysninger, og annen potensielt sensitiv data, kritisk.
Prosjektet deles inn i fire overordnede temafokus/delprosjekt:
• P1: Anonymisering- og pseudonymiseringsmekanismer for sensitive KI-applikasjoner.
• P2: Anvendt kryptografi for beskyttelse av sensitive data i hele behandlingskjeden av KI-applikasjoner.
• P3: Personvernbevarende trening av KI med datasett som inneholder personopplysninger/sensitive data.
• P4: Case-studier rettet mot KI-applikasjoner i helse-, justis- og forsvar.
• P5: PhD-avhandling.