Efficient and Secure Over-the-Air Machine Learning for 5G and Beyond

• Applicant: Professor Dr.-Ing. Slawomir Stanczak
• Subject Area: Communication Technology and Networks, High-Frequency Technology and Photonic Systems, Signal Processing and Machine Learning for Information Technology
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 467514322

In current 5G networks, Machine Learning (ML) plays an ever more important role, and it is expected to become even more essential in the future. The envisioned uses of ML are extremely diverse and range from tasks that are necessary for operating the communication infrastructure itself, such as Quality-of-Service prediction, anomaly detection and channel estimation to a wide range of applications such as e-health, smart manufacturing and sensor network-aided early warning systems for natural disasters. OTA ML (Over-The-Air Machine Learning) has emerged as a hot new research topic over the last few years. Many of the proposed techniques draw on the Federated Learning (FL) framework which first appeared in the context of distributed computing, where for various reasons such as privacy concerns the full data cannot be exchanged between communication partners, but the communication itself follows the classical paradigm of source-channel separation. These new OTA ML schemes treat a large variety of scenarios of Horizontal FL where every agent in the system observes the same set of features, but the training samples are available in distributed form only. In many envisioned applications, however, also the case of Vertical FL is of interest where the agents in the system can observe different features, and many basic questions in this direction are still open, some of which we intend to address in this project. In many envisioned application scenarios of OTA ML (e.g., e-health and smart manufacturing), security and privacy are major concerns. However, integrating state-of-the-art cryptography or physical layer security with existing OTA ML approaches is not straightforward at all, and we expect that the only way to make them secure is to design them from the ground up with the security concerns in mind. As a first step on this path, within this project we intend to develop techniques to protect OTA ML schemes against passive eavesdropping attacks. This will require a substantial amount of fundamental research and some of this research can also be expected to have an impact in areas of Physical Layer Security that do not necessarily deal with OTA ML. Specifically in this context we plan to investigate deployment of meta-surfaces that help shape the wireless channel in such a way that protection against these eavesdropping attacks is possible. The project we propose is therefore a blend of fundamental and applied research. Some of it, such as the Vertical FL schemes we are going to design, will be demonstrated in simulations and even hardware demonstrators, while we expect a large part of our research on security in the OTA ML context to be on a fundamental level. This means that the path towards technical applicability will have to be continued in successive research projects, but also that it can have an impact that goes beyond the immediate use in conjunction with OTA ML techniques.

• DFG Programme: Research Grants