Their work, published in a research paper titled "GNN-based Deep Reinforcement Learning with Adversarial Training for Robust Optimisation of Modern Tactical Communication Systems," uses a hybrid Machine Learning approach based on graph neural networks and deep reinforcement learning.

Loevenich, the main author and PhD student, explains the solution: "In our research, we focus on optimising modern tactical communication systems in mobile ad hoc networks, which are especially critical in military contexts. We address two main problems: predicting fast-changing network conditions, which we do using graph neural networks (GNNs), and dynamically configuring the system to ensure robust communication, which is improved using deep reinforcement learning (DRL). GNNs allow us to effectively model and predict the complex topology and dynamically changing status of the network, while DRL develops adaptive strategies to optimise network performance under those conditions."

Improved robustness through adversarial training

The researchers drew attention to a key part of the work in the implementation of adversarial training, which substantially improves the robustness and security of the developed models. This approach primes the system to recognise faults and potential attacks and combat them effectively. Loevenich continues: "The research and development work that was part of my doctorate at Fraunhofer FKIE, and especially the development of the DRL components and the integration of adversarial training, contribute significantly to the performance and reliability of tactical communication systems, improving operational effectiveness in military fields of application."

Integration into different tactical communication products

The researchers are currently working on integrating the developed models into tactical communication products of different model sizes and with different hardware requirements. "A key challenge is to develop an ML pipeline to ensure that the models function reliably both online and offline," Loevenich explains. "We are focusing on how different instances of these models, which can perform different functions, can exchange information effectively. This requires carefully adjusting the system architecture and algorithms for optimum performance and communication between the models, while also ensuring that the system meets the requirements in terms of hardware efficiency and adaptability to different practical scenarios."

Promoting innovations in the defence sector

The presentation of the "EDA Research, Technology, and Innovation Papers Awards" is part of the EDA’s strategy to engage in the field of research and development to gain insight into technological trends and promote cutting-edge innovations in the defence sector. Dr Markus Antweiler, deputy institute director of Fraunhofer FKIE, says: "The award is an important acknowledgement of the research team’s years of work to overcome technological limits with artificial intelligence processes, and to close ability gaps in security and defence. We offer our heartfelt congratulations."