SpliDT: Partitioned Decision Trees for Scalable Stateful Inference at Line Rate#

Murayyiam Parvez et al. (Purdue, Michigan, UCSB, Open U. Israel, NIKSUN)

Why: In-network ML for security monitoring usually assumes all features upfront; this confronts the data-plane stateful-feature constraint honestly.

What: A framework supporting up to 5x more stateful features and scaling to millions of flows by partitioning decision-tree inference over a sliding packet window.

How: Sequential subtrees each using their own top-k features, with an in-band control channel via packet recirculation managing transitions and reusing match keys/registers; a joint training/DSE workflow optimizes feature allocation, depth, and partitioning.

Where applicable: RMT programmable switches/SmartNICs; a general data-plane-ML technique constrained by switch resources.