Multi-task Aware Resource Efficient Traffic Classification via in-Network Inference

With the advancement of in-network intelligence (INI) capabilities in programmable data planes (PDP), there is a rising demand for efficiently executing multiple tasks within the constraints of programmable switches. However, reliance on single-task learning (STL) models for INI faces limitations in meeting this demand. To address these challenges, we develop a multi-task aware resource efficient traffic classification via in-network inference (MARTINI) scheme. MARTINI is a multi-task learning (MTL) approach that utilizes a binary neural network (BNN) architecture, where multiple tasks share the same hidden layers, enabling efficient parameter sharing and reducing resource consumption on programmable switches without substantial degradation in classification performance. We implemented MARTINI on the BMv2 software switch, showing that it reduces memory usage by up to 48% and shortens inference processing time by 40% compared to the STL model, while maintaining sufficiently high classification performance. Furthermore, it demonstrates that performance can be maintained regardless of model complexity in terms of the number of tasks and classes.