In-Transit Payload Fingerprinting for High-Throughput Data Streams

Abstract

Programmable network data planes enable limited in-network computation, but most prior work has focused on narrow primitives or small-scale prototypes. This paper demonstrates real-time processing of high-dimensional image streams at near-terabit rates on the FABRIC testbed using programmable switches. We design a lightweight rare phenomena detection mechanism that computes payload-consistent fingerprints using a hybrid neutralized-TCP checksum technique compatible with P4 switch pipeline constraints. The system is implemented on an Intel Tofino switch and evaluated across the FABRIC terabit core ring using distributed traffic generators. Our results demonstrate reliable detection of payload changes and sustained 755 Gbps in-network fingerprinting throughput. To our knowledge, this is the first demonstration of payload-consistent fingerprinting implemented entirely within a programmable switch pipeline and evaluated at multi-100 Gbps scale on a wide-area experimental infrastructure.