DA-DCC-IDS: Trustworthy Open-Set IoT Intrusion Detection Under Temporal and Device Uncertainty

Abstract

Internet of Things (IoT) intrusion detection on a gateway has to handle three coupled risks: heterogeneous device traffic distributions, unknown attack families arriving at inference time, and softmax confidence that does not separate correct predictions from confident errors on novel traffic. We present DA-DCC-IDS, a pipeline whose central contribution is an open-set rejection layer that survives temporal-feature ablation: a per-device Mahalanobis density fitted on a 150-dimensional truncated-SVD projection of LightGBM leaf indicators, fused with six auxiliary out-of-distribution signals through a seven-weight convex combination whose weights follow from per-signal tuning AUROC. The closed-set track stacks LightGBM, XGBoost, and CatBoost under a logistic-regression meta-learner with per-class thresholds; it is honest about depending on runtime-observable burst dynamics. We retrain the full pipeline on TON_IoT, IoT-23, and N-BaIoT under burst- or capture-grouped splits and benchmark five deep-learning baselines (MLP, 1D-CNN, LSTM, autoencoder, Transformer) trained under the same partition, scaler, and epoch budget. The open-set fusion reaches AUROC 0.961 on TON_IoT, 0.973 on IoT-23, and 0.946 on N-BaIoT held-out attack families. We fit the fusion weights on validation-side known positives together with 5,044 labelled unknown tuning rows, which makes the primary protocol a labelled-OOD tuning setup rather than zero-shot open-set recognition; no test-side known row enters the weight fit. We report the zero-shot case as a separate leave-one-family-out test that fits the weights on one unknown family and evaluates on a second family the weight fit never sees: TON_IoT AUROC reaches 0.972 when we tune on XSS and evaluate on scanning, and 0.952 when we reverse the two. The framework is device-manifold-aware and does not require a trustworthy runtime device label; a truly unseen device type falls back to near-chance detection, which we characterise in Section V-J. Paired McNemar tests on closed-set predictions and paired-bootstrap tests on AUROC place DA-DCC-IDS above every deep-learning baseline at $p \lt 0.001$ in all 22 paired comparisons on IoT-23 and N-BaIoT after cross-dataset replication under retraining. The closed-set macro F1 of 0.867 on TON_IoT drops to 0.330 when temporal and burst features are removed, while open-set AUROC stays at 0.961 under the same removal. The geometric reject layer is decoupled from temporal predictiveness; the closed-set head is not. We report bundle size 46.5 MB, $53.3~\mu $ s per-sample latency on a commodity CPU baseline, and the limits of an unseen-device case (AUROC near chance).