Two talks at KI@Bw26
22 Mai 2026
From the 19th to the 21st of May, the fourth KI@Bw conference took place in Hamburg, Germany, where current developments, challenges and opportunities for the use of artificial intelligence within the Bundeswehr were the main topic of discussion. The chair of Data Science was represented by Moritz Hennen, Yeong Su Lee and Hendrik Bothe. The first day of the conference was dominated by discussions regarding ethical challenges that surface with the introduction of artificial intelligence, both within the context of the Bundeswehr but also in the broader societal scope. This perspective was further enlightened by a talk that approached these problems from a theological point of view. The following two days provided a diverse, three-track program with many different topics, including but not limited to satellite, sonar and audio data processing and information gathering, drone usage and software development in the age of artificial intelligence. Furthermore, two hands-on workshops allowed participants to gather practical experience with agent- and RAG-driven information gathering tools.
Moritz Hennen presented his talk on “Information gathering with OSINT: Scalable methods for the automatic modeling of knowledge graphs” and introduced scalable and efficient methods for large-scale transformation of unstructured data to a knowledge graph, as an alternative to the application of large language models (LLMs) for the same task. Motivated by achieving high throughput whilst significantly reducing inference time, a method developed at our chair, ITER [1], was introduced. The talk closed with an appeal to all attendees as to not blindly rely on the usage of language models for solving tasks where conventional approaches are just as applicable, particularly with regards to their energy usage.
[1] https://aclanthology.org/2024.findings-emnlp.655/
Yeong Su Lee and Hendrik Bothe presented the talk titled Robust Obfuscation to Protect Sensitive Image Data on Tuesday. The center of their talk was the question on how profile pictures in social networks may be protected, in a way that prevents the re-identification of their profile owners whilst allowing the data to be used for further analysis. The introduced method does not directly change facial features in the profile pictures, but extracts demographic and semantic features and generates new profile pictures based on those features using image generation models. The method attempts to preserve important facial features whilst ensuring anonymization. The talk closed with a live demonstration to highlight the potential of this approach.
Image sources: Moritz Hennen (RI CODE)
