AgriEdge: a leakage-free precision-agriculture IIoT intrusion detection benchmark

A leakage-free precision-agriculture IIoT intrusion detection benchmark, rebuilt from the raw per-device captures of Edge-IIoTset under uniform parsing. Why this exists: Edge-IIoTset's curated subsets carry a serialisation artifact. Normal-traffic and attack captures were parsed separately, so the placeholder written for an absent protocol field is spelled 0 in one branch and 0.0 in the other. Four of the seven columns the distributed preprocessing recipe tells researchers to one-hot encode therefore recover the attack/normal label with accuracy 1.0000 on their own, encoding file provenance rather than network behaviour. This benchmark is rebuilt from the same raw captures with a single parser and a single placeholder token used everywhere. After the rebuild no column separates the classes above a rate of 0.0288, and the strongest single-column accuracy is 0.6179. Shape: 1,276,122 rows by 53 columns. 49 protocol feature columns, all stored as strings; Attack_label (0 = normal, 751,559 rows; 1 = attack, 524,563 rows); Attack_type (15 values); source_device (19 values); source_layer (perception 601,563, actuation 149,996, adversary 524,563). The feature columns are stored as text deliberately. Casting to float destroys the evidence, because pandas coerces the strings 0 and 0.0 to the same value, which is the most likely reason the original artifact went unreported. Intended use: device attribution is retained so that leave-one-device-out evaluation is possible. Random stratified splitting still overstates generalisation here, because packets from the same device are highly correlated. Transfer between perception sensors is nearly free, but collapses across the perception/actuation boundary: random forest falls from 0.9988 to 0.5083 balanced accuracy when the Modbus gateway is withheld. Please report leave-one-device-out or grouped splits alongside any random-split number. The file is regenerable by running experiments/03_build_agribench.py against a local copy of Edge-IIoTset. Code, experiment scripts and result tables: https://github.com/MostafaGalal1/agriedge (DOI 10.5281/zenodo.21941210) Derived from Edge-IIoTset by Ferrag, Friha, Hamouda, Maglaras and Janicke (IEEE DataPort, doi:10.21227/mbc1-1h68; described in IEEE Access 10:40281-40306, 2022, doi:10.1109/ACCESS.2022.3165809). Free for academic use per its own licence.