Petr Taborsky: The project will create a method that provides for new services creation in an upcoming interconnected word of IoT and 5G networks without compromising end user privacy. It will build on a research in fields of federated machine learning, conditional density estimators leveraging deep neural networks, Bayesian filtering, generative models and adversarial machine learning.
As homes get smart, cars become connected and factories develop ever more automation, the need for connectivity and privacy become paramount. And it is necessary to maintain user confidence in connected devices as they become more integrated into our lives.
Machine learning based on deep neural network architectures have achieved remarkable results in a wide variety of domains. The most profound and developed deep network applications solve perceptual AI problems such as objects recognition in images or speech recognition based on sound signals. Applications of deep learning on connectivity (a.k.a IoT) data, including location, behavioral and health data streams from mobile devices, mobile phones and likes, are less common and developed.
In most communication networks and in telecommunication ones in particular, connected devices spawn "meta data" as inevitable product of request processing i.e. call number +4578987xx or download http://...
In other words, no matter how protected data is on the device, meta data, timestamps, location, destination sockets and alikes, being linked to IP address & port, MSISDN but also ?soft? cookie-based identifiers, proliferate, carrying insight into "privacy" behind connected device.
Latest development in federated machine learning shows a way to protect privacy ?on the device?. Nevertheless and as far as network meta data goes, further enhancements and novel methods of privacy protection going beyond current measures of limited access and encryption are needed to unlock full potential of this data for research and new services.
Set of connected devices can be viewed as a time-varying system of sensors which is indirectly observed in network through (noisy) measurements. The ?state of this system? refers then to the collection of dynamic variables such as location, online activity as well as descriptive dimensions i.e. age, which fully describe the system.
In this project we will design a dynamic density model for ?state of the system? of connected devices that optimizes end-to-end data processing pipeline in order to maximize user privacy protection in such networks while providing data for research and (new) services.
It will be based on deep neural networks (DNN) and other methods i.e. Hierarchical Dirichlet Process (HDP) as conditional density estimators (CDE) and novelty detectors, that are going to be trained federally and updated dynamically on an ongoing basis, leveraging Bayesian filtering and generative models.
PhD project title: Federated deep learning for privacy preserving mobile data modelling
Effective start/end date 01/01/2018 → 31/12/2020
Main supervisor: Lars Kai Hansen, Co-supervisor: Finn Årup Nielsen
Research section: Cognitive Systems