The 16th International Conference on Mobility, Sensing and Networking (MSN 2020)
Keynote Talks
Session Keynote-1
Keynote I
Conference
5:00 PM
—
6:00 PM JST
Local
Dec 18 Fri, 3:00 AM
—
4:00 AM EST
Scaling the Edge
Prof. Jon Crowcroft (University of Cambridge, UK)
1
Extreme federation of processing data in mobile networked sensor systems is becoming mainstream as an approach. It reduces load on the uplinks, it saves energy and potentially provides better privacy for personal data.
There are a variety of techniques ranging from simple aggregation, compressive sensing, and edge-machine learning, where models are locally acquired, and model parameters are distributed, so nodes can further refine their models.
There are a number of challenges to scaling such approaches. Firstly to scale federated learning to billions of nodes needs some way to scale even just sharing model parameters - I will discuss some of these, including sampling of model parameters (thinning, probabilistic update) and self organising hierarchies of aggregation (model parameter servers). For some Machine Learning algorithms, there may be updates from the federated model back to nodes to adjust their learning (e.g. regret) as well.
Some schemes may require synchronisation of learning steps. All these need to scale out, and techniques form data centers may, surprisingly be applicable, even though we are often in a much less rich networking environment, even without full connectivity or symmetric bandwidth or reachability.
Federation alone is not a complete solution to privacy, and there are some further techniques may be needed to reduce the loss of confidentiality - e.g. differential privacy is useful, but also more fundamental approaches such as secure multi-party computation, in extreme cases.
Secondly, there is the problem of bad actors injecting false data (pollution). Then there is the omnipresent presence of possible DDoS attacks.
Thirdly, a federated model may present some challenges to model explain-ability or interpret-ability. There are interesting trade-offs between these requirements, and those of privacy.
There are a variety of techniques ranging from simple aggregation, compressive sensing, and edge-machine learning, where models are locally acquired, and model parameters are distributed, so nodes can further refine their models.
There are a number of challenges to scaling such approaches. Firstly to scale federated learning to billions of nodes needs some way to scale even just sharing model parameters - I will discuss some of these, including sampling of model parameters (thinning, probabilistic update) and self organising hierarchies of aggregation (model parameter servers). For some Machine Learning algorithms, there may be updates from the federated model back to nodes to adjust their learning (e.g. regret) as well.
Some schemes may require synchronisation of learning steps. All these need to scale out, and techniques form data centers may, surprisingly be applicable, even though we are often in a much less rich networking environment, even without full connectivity or symmetric bandwidth or reachability.
Federation alone is not a complete solution to privacy, and there are some further techniques may be needed to reduce the loss of confidentiality - e.g. differential privacy is useful, but also more fundamental approaches such as secure multi-party computation, in extreme cases.
Secondly, there is the problem of bad actors injecting false data (pollution). Then there is the omnipresent presence of possible DDoS attacks.
Thirdly, a federated model may present some challenges to model explain-ability or interpret-ability. There are interesting trade-offs between these requirements, and those of privacy.
Session Chair
Lars Wolf (Technische Universitaet Braunschweig, Germany)
Made with in Toronto · Privacy Policy · © 2022 Duetone Corp.