Subproject "Software Platform for Shared AI"

A software platform for devloping and testing shared AI is envisoned that goes beyond standard federated learning by taking transfer learning into account:

• Development Platform: to support the configuration, instantiation and orchestration of pipelines for training transfer learning and confidentiality preservation algorithms;
• Test Environment: for modeling threat and attack scenarios to evaluate privacy and confidentiality protection and identify vulnerabilities;

Collaborative AI between distributed parties relies on the transmission and sharing of data, parameters, and model fragments to improve the performance of an aggregated model compared to a single model. However, such a setting poses vulnerabilities that can be exploited by adversarial users. To cope with related security threats we have been developing a tool ecosystem consisting of:

1. Design methodology for hyper-parameter tuning in domain adaptation setup to minimize performance loss
2. Methodology for semi-supervised ensemble learning to enhance malware detection accuracy and privacy at the same time
3. Analysis tool addressing poisoning attack effects by utilizing backdoor learning curves
4. Information-theoretic analysis and design tool to tune the tradeoff between privacy leakage and data utility

These approaches have brought about follow-up research projects in industry 4.0 as well as German KI-SIGS project with Uni Lübeck as coordinator; KI-SIGS is dedicated to the development of an "AI Space for Intelligent Health Systems" in collaboration of northern German AI institutes in Bremen, Hamburg and Schleswig-Holstein together with medical technology companies and partners of the university hospitals.