Date and LocationTuesday June 04, 2019 9:00am
Recent efforts highlight the promise of data-driven network design. Rather than using manually configured rules, a data-driven design leverages real-life observations to make decisions. However, data-driven network design faces the challenges of data complexity and scalability. Inherent complexities in network behavior, combined with artifacts in network measurements (e.g., noise, bias, error and even malicious inputs), often lead to poor design decisions that degrade network performances.
My research focuses on addressing the above challenges using machine learning. In this defense talk, I will present two of my recent works. First, I will show how we build building deep neural network (DNN) models to capture spectrum usage patterns and use them as baselines to detect spectrum usage anomalies resulting from faults and misuse. In particular, our work addresses the problem of model scalability, such that the DNN models can be easily deployed on any observers (mobile and static). All the observers in a single cellular cell will run the same model regardless of their context, and will only switch to a different model when they move to a different cell. We also develop efficient “transfer learning” based training to quickly adapt models using a small amount of local spectrum measurements. Second, I will briefly introduce my recent work that looks at the weakness of current ML-based designs. We show that adversarial jamming via ultrasonic signals can easily break existing speech recognition systems (designed to be robust to environmental noises).