Work

Decision-Level versus Feature-Level Novelty Detection for Open-Set Intrusion Detection: A Cascaded EfficientNet--GBDT and Isolation Forest Framework Evaluated on TON_IoT

Year: 2026
Type: preprint
Abstract: Closed-set classifiers on Internet of Things (IoT) and Industrial IoT (IIoT) intrusion benchmarks now reach accuracies at which aggregate accuracy no longer separates competing methods, while the capability that matters most for critical infrastructure, namely sensible behaviour on attacks never labelled during training, remains weakly addressed. This study proposes and evaluates a cascaded open-set intrusion detector on the network component of the Telemetry, Operating System and Network IoT (TON\_IoT) dataset. Flow records are reduced by Principal Component Analysis (PCA) and Electromagnetic Field Optimisation (EFO), a compact EfficientNet extracts a learned embedding, a Gradient-Boosted Decision Tree (GBDT) ensemble classifies that embedding, and flows whose maximum GBDT class probability falls below 0.95 are forwarded to an Isolation Forest for novelty detection. Injection and Man-in-the-Middle (MITM) traffic are withheld from training to measure open-set behaviour. The study makes four contributions: (i) a controlled comparison of attaching the Isolation Forest to the reduced input features (decision-level) or to the learned embedding (feature-level) under identical preprocessing, data partitioning and withheld-attack protocol; (ii) evidence that replacing the softmax head with a GBDT head leaves closed-set predictions unchanged; (iii) operating points reported as alarm precision at stated attack base rates rather than detection rate alone; and (iv) measured training, storage, latency, memory and energy costs. On the eight known classes, the GBDT head attains 98.26\% accuracy (macro-F1 0.9789) on 25,421 test flows, identical flow for flow to the EfficientNet softmax baseline. On withheld attacks, the classifier alone flags none as unknown, decision-level attachment identifies 81.4\% and feature-level attachment 76.1\%, indicating that supervised training compresses away the variation that makes unseen traffic unusual. Withheld-attack capture rises from 76.1\% at a 0.65\% benign false-alarm rate (54.2\% alarm precision at a 1\% attack base rate) to 99.8\% at 13.9\%, where alarm precision falls to 6.8\%. The cascade costs 4.7 times the inference time of the feature extractor alone and occupies 91.4 MB. These findings show that the placement of the novelty detector, not the classification head, determines open-set capability, with direct implications for intrusion detection at petroleum-pipeline monitoring gateways.
Language: English
Open Access status: green
OpenAlex ID: W7236795070
Cites: 31
Cited by:
Related to: 10