Analysing Foundation models in the security context

Dr Daniel Gille, Head of Artificial Intelligence and Key Technologies at Cyberagentur. Photo: Andreas Stedtler/Cyberagentur

Information and networking at the intersection of AI and geodata

Addressing the scheduled research project „Holistic Evaluation of Generative Foundation Models in a Security Context“ (HEGEMON), the Agentur für Innovation in der Cybersicherheit GmbH (Cyberagentur) is organising a virtual partnering event on 31 July 2024. The event offers a platform for potential bidders and interested parties to learn about requirements and opportunities of the research project, and to exchange ideas.

The term “foundation model” describes complex neural networks that have been trained on large, unspecific amounts of data, and are characterised by their versatility in a wide range of applications. Foundation models enable the generation of complex text, image or audio content based on free input (prompts). However, these models require a high initial training effort and comprehensive evaluation, particularly in the context of internal and external security.

With this in mind, the research team led by Dr. Daniel Gille, Head of Artificial Intelligence in Cyberagentur’s Key Technologies department, has developed various research assignments and use cases. Data sets provided by the Bundeswehr Geoinformation Centre will also be a part of the planned tender. „The research questions are divided into three main areas,“ explains Daniel Gille: „Benchmarking as a problem of interdisciplinary basic research, development of a holistic foundation model benchmark against the application background of geoinformation, and development of a demonstrator for solving specific tasks in geoinformation.“

The development of benchmarks and adaptation of foundation models as well as their demonstrator implementation are to be realised in a unique competitive constellation in which each participant is in direct comparison with all other participants in terms of both benchmark and model development. Each model is evaluated and ranked against all benchmarks developed. All benchmarks and further research results are also evaluated separately regarding their specific contribution.

The virtual partnering event on 31 July 2024, from 13:30 to 15:30 CEST, offers the opportunity to learn more about this research project, and to exchange ideas with potential partners. Further information on the event and registration can be found on the research project website.

Further information:

German Version