An IMDEA Networks research wins ‘Best Demo Award’ at IEEE NetSoft 2023
The winning article is "Showcasing In-Switch Machine Learning Inference"
27 June 2023
The paper “Showcasing In-Switch Machine Learning Inference” by Aristide Tanyi-Jong Akem, Beyza Bütün, Michele Gucciardo, Marco Fiore, has received the award in the “Best Demo Award” category of the IEEE NetSoft 2023 conference held in Madrid on June 19-23.
The study carried out by this IMDEA Networks research team was the first to implement and validate flow-level classification in production programmable switches. “It is therefore a step forward towards moving into the network hardware the classification functionalities that are traditionally realized in software in the control plane. This has huge advantages in terms of latency, which is reduced from milliseconds in the control plane to less than 100 nanoseconds in our tests,” explains Marco Fiore, Research Professor at the institute and one of the authors of the award-winning paper.
In this paper they have demonstrated how Random Forest models can be deployed in modern programmable switches in order to classify network traffic flows, i.e., entire sequences of packets between two hosts belonging to the same exchange (e.g., a same video streaming session or file download).
In this way, they have made progress with respect to the work already done so far, which only classifies each package separately. Indeed, operating at flow level gives access to new features (e.g., inter-arrival time between subsequent packets of the same flow) that are not available when only looking at packets individually.
“Ultimately, a flow-level approach results in much more precise classification: we achieve 99% accuracy in the service classification use case we demonstrated, with an 8 percent point gain over previous packet-level solutions”, adds Fiore.
The demonstrator is the result of a successful line of research that has led to works published at IEEE INFOCOM 2023, IEEE MetaCom 2023 and the NativeNI workshop co-located with ACM CoNEXT 2023.
Source(s): IMDEA Networks Institute