At the physical layer of networks – where radio signals or cables carry digital information – Gansekoele developed a new generation of neural receivers. These receivers can handle different types of signals more flexibly. This is important because networks increasingly have to serve a wide range of devices and applications simultaneously, each with its own demands for speed and reliability. A flexible receiver makes the network more robust and efficient.
He also investigated how deep learning can help recognize sounds underwater, such as distinguishing between ship engines, whales, and natural noise. Such techniques can be used both to map disturbances to marine life and to develop more reliable underwater communication, for example for sensors or robotics.
Vulnerabilities in videostreaming
But where deep learning can strengthen communication, it can also reveal weaknesses. In the case of video streaming, Gansekoele found that current security is less watertight than often assumed. Videos on platforms such as YouTube are normally encrypted and transmitted in small segments. Yet he discovered that the way these segments arrive reveals recognizable patterns.
With a deep learning model, a third party – such as an internet provider – can infer which video someone is watching, without breaking the encryption. Previously, this was only known in theory, but the new approach makes it much easier and more scalable. This shows that protocols like HTTPS are not fully resistant to sophisticated analytical methods.
Gansekoele emphasises that this does not mean users are directly unsafe, but it does show that protocols such as MPEG-DASH and HTTPS need to be strengthened against such analyses. His recommendations range from small technical adjustments in video encoding to more structural improvements in the standards themselves.
Privacy and Fairness
Finally, Gansekoele studied federated learning, a method in which different organisations collaboratively train an AI model without sharing their data. This is particularly relevant for sensitive applications such as facial recognition. He shows that this approach especially benefits weaker parties – organisations with smaller or more diverse datasets – helping them achieve better performance and thereby ensuring fairer outcomes.
Next Steps
The findings make clear that deep learning offers both opportunities and risks for the communication systems of the future. In particular, the video streaming research highlights the urgency of making existing protocols more resilient against deep learning methods that can still identify encrypted data traffic.
Picture: Shutterstock
