We interviewed all three award winners and asked them about their motivation, the difficulties and prospects of their research work. Read the full interviews in our online magazine Fraunhofer InnoVisions (in German only).

Rezaul Karim from Fraunhofer FIT won first prize with his work "Interpreting Black-Box Machine Learning Models with Decision Rules and Knowledge Graph Reasoning".

"It is well known that decision-making processes in machine learning models are not comprehensible and the whole thing ends up in a black box. This can be problematic in many cases. Data distortions or incorrect training data can produce undesirable or even discriminatory behavior in AI systems. This can be problematic, especially in areas such as medicine, where artificial intelligence has a significant impact on people's lives. It is therefore all the more important, especially in view of the growing use of AI in all areas of society, to make predictions and decisions comprehensible."

His work aims to improve the interpretability and explainability of opaque machine learning models without significantly compromising prediction accuracy.

Franziska Boenisch from Fraunhofer AISEC won second prize with her dissertation "Secure and Private Machine Learning".

"Machine learning models should work correctly and not reveal too much information about their sensitive training data. However, we find that developers have a particularly low awareness of privacy in machine learning and rely on third-party services for their implementation. This often leads to privacy risks. Our research shows that federated learning in particular poses great risks to users' private data, as they can leak data directly to companies."

In her dissertation, Franziska Boenisch examines security and data protection aspects in the field of machine learning. She identifies attack vectors and security vulnerabilities and emphasizes the need to design ML methods in a secure and privacy-friendly way.

Stylianos Ioannis Mimilakis from Fraunhofer IDMT was awarded third prize in 2023 for his dissertation "Deep Learning-Based Music Source Separation".

"The dissertation is basically about data-driven methods for separating musical instruments and in particular groups of musical instruments from a digital music recording. In a karaoke system, for example, the vocals can be separated and the background music output. Other applications could be the revival of historical recordings that require a special separation of the music sources."

Mimilakis' work examines the problem of music source separation using deep learning methods. The three aspects of signal processing, neural architecture and signal representation are considered. He is investigating new algorithms and network architectures for filtering and separating music sources.