Electrical Engineering is a core competence reflected in all our research groups. We research and develop innovative circuits for Swiss companies in ultra-low-power electronics, with or without energy harvesting; wireless communication systems; high performance computing equipment and safe and dependable systems. We also research and develop innovative systems in VHDL on FPGAs and software for high performance signal processing, network and communication technologies and embedded machine learning.

Computer Science for Embedded Systems is a core competence reflected in all our research groups. We research and develop software for both IoT/ultra-low-power systems as well as high performance networking and signal processing. We place special focus on high-quality software engineering, applied security, attachment and interaction between embedded (IoT) devices and data-bases and the cloud, signal processing, (embedded) operating systems and embedded machine learning. Our applied research output is used by industry in ultra-low-power systems, signal processing, communication, both high performance and wireless, embedded machine learning and safe and dependable systems.

Data Science in Embedded Systems is a competence reflected in our IoT and embedded machine learning groups. We utilise data-mining algorithms in video processing for safe and dependable systems and in IoT systems in pattern matching whilst the embedded machine learning groups rely heavily on data science technologies for the optimisation of machine learning for resource restrained systems.

Information and Cyber Security

How can device manufacturers protect their low power IoT devices?
The thesis proposes a systematic design flow for secure IoT devices. Furthermore, it provides a proof-of-concept by implementing a real-world example using a Secure Microcontroller with a Trusted Execution Environment (ARM TrustZone). The result demonstrates state-of-the-art embedded security measures and their adoption in a secure, cloud-based application.

The goal of this thesis was to investigate methods to make neural networks adaptive on microcontrollers. In this context, adaptive means, that the neural network on the microcontroller can automatically adapt to the specific environment. This brings numerous advantages, because often the environment where microcontrollers are used is not static and for example measured sensor data can drift over time. The important step was to analyze algorithms which allow neural networks to adapt to changing environments. This process is known as ”domain adaptation”, where the network was originally trained on different data than in the real application. Different algorithms were investigated to see if they can be adapted to run on microcontrollers. The most promising algorithm was adapted for the use on a microcontroller and evaluated with different data sets, and compared to the original algorithm. The adaptation was successful and delivered comparable or better results than the original algorithm.