LEAKAGE-AWARE LATENT-DIMENSION SELECTION FOR INTRUSION DETECTION IN AGRICULTURAL IOT NETWORKS

Agricultural Internet of Things (AG-IoT) intrusion detection requires models that limit representation complexity without weakening the credibility of their evaluation. This study investigates whether a learned intrusion-detection representation can be reduced to a sparse latent subset while retaining a useful clean-detection profile. A compact one-dimensional CNN encoder, referred to as CNN-Lite, learns a 128-dimensional latent representation; IWHO-Lite performs binary, sparsity-aware dimension selection; and a Random Forest provides the terminal classification score. Data-dependent preprocessing is confined to the training scope, while model-development decisions remain isolated from score calibration, operating-threshold selection, and final testing. Exact-vector overlaps are additionally audited before evaluation. Across Farm-Flow, CICIoT2023, and UNSW-NB15 using three independent model seeds, IWHO-Lite retained 63.00, 37.67, and 41.67 latent dimensions, corresponding to compression rates of 50.78%, 70.57%, and 67.45%, respectively. Mean clean-test F1-scores were 0.4724, 0.9953, and 0.7480. Mean batch-inference latency ranged from 0.0973 to 0.1344 ms per sample in the recorded experimental environment. Farm-Flow combined high recall with a substantial false-positive burden, highlighting strong dataset dependence. The results support a leakage-aware performance-parsimony characterization of latent-dimension selection rather than universal predictive superiority or hardware-validated efficiency.