IoT botnet detection has advanced, yet most published systems are validated on a single dataset and rarely generalise across environments. Heterogeneous feature spaces make multi-dataset training practically impossible without discarding semantic interpretability or introducing data integrity violations. No prior work has addressed both problems with a formally specified, reproducible methodology. This paper does. We introduce BRIDGE (Benchmark Reference for IoT Domain Generalisation Evaluation), the first formally specified heterogeneous multi-dataset benchmark for IoT intrusion detection, unifying CICIDS-2017, CIC-IoT-2023, Bot-IoT, Edge-IIoTset, and N-BaIoT through a 46-feature semantic canonical vocabulary grounded in CICFlowMeter nomenclature, with genuine-equivalence-only feature mapping, explicit zero-filling, and per-dataset coverage from 15% to 93%. A leave-one-dataset-out (LODO) protocol makes the generalisation gap precisely measurable: all five evaluated architectures achieve mean LODO F1 between 0.39 and 0.47, and we establish the first community generalisation baseline at mean LODO F1 = 0.5577, a result that shifts the agenda from single-benchmark optimisation toward cross-environment generalisation. We propose TCH-Net, a multi-branch network fusing a three-path Temporal branch (residual convolutional-BiGRU, stride-downsampled BiGRU, pre-LayerNorm Transformer), a provenance-conditioned Contextual branch, and a Statistical branch via Cross-Branch Gated Attention Fusion (CB-GAF) with learnable sigmoid gates for dynamic feature-wise mixing. Across five random seeds, TCH-Net achieves F1 = 0.8296 +/- 0.0028, AUC = 0.9380 +/- 0.0025, and MCC = 0.6972 +/- 0.0056, outperforming all twelve baselines (p < 0.05, Wilcoxon) and recording the highest LODO F1 overall. BRIDGE and the full pipeline are at https://github.com/Ammar-ss/TCH-Net.
Figures & tables
Method
Multi- branch
Gated Fusion
Multi- dataset
IoT Focus
Full Ablation
RF / XGB [ 8 , 9 ]
—
—
—
—
—
BiLSTM [ 17 ]
—
—
—
✓
—
Transformer [ 18 ]
—
—
—
✓
—
Kitsune [ 20 ]
✓
—
—
✓
—
N-BaIoT [ 19 ]
—
—
—
✓
—
DeepDefense [ 21 ]
—
—
—
✓
—
Table 1: Qualitative comparison with representative prior work. (✓) = supported; (—) = not supported.
Dataset
Capture Tool
Year
Coverage
Tier
CICIDS-2017
CICFlowMeter
2017
93%
Primary
CIC-IoT-2023
CICFlowMeter
2023
87%
Primary
Bot-IoT
Argus
2019
39%
Primary
Edge-IIoTset ⋆
Wireshark
2022
22%
Supplementary
N-BaIoT ⋆
Kitsune
2018
15%
Supplementary
Table 2: Overview of the five BRIDGE datasets. ⋆ = Supplementary; low coverage reflects non-flow-level capture.
Group
Semantic Category
Indices
Count
1
Flow rates, durations, pkt/byte counts
0–16
17
2
Packet size & IAT statistics
17–37
21
3
TCP flag indicators
38–43
6
4
Header length & window size
44–45
2
Total
46
Table 3: 46-feature canonical vocabulary by semantic group.
Figure 1: Feature coverage of the 46-feature canonical vocabulary across five BRIDGE datasets. Blue cells indicate a genuinely matched feature; grey cells indicate explicit zero-fill (feature absent from that dataset). Column groups correspond to the four semantic categories in Table 3 . Coverage percentages are reported in Table 4 .
Dataset
Benign
Attack
Total
Atk%
CICIDS-2017
19,321
14,350
33,671
42.6%
CIC-IoT-2023
3,964
3,001
6,965
43.1%
Bot-IoT
22
16
38
42.1%
Edge-IIoTset
30,951
23,435
54,386
43.1%
N-BaIoT
13,557
10,055
23,612
42.6%
Combined
67,815
50,857
118,672
42.9%
Table 5: Post-balancing record counts per dataset.
Figure 2: Full TCH-Net architecture overview across five zones: inputs, shared feature projection, three parallel branches (T, C, H), CB-GAF fusion, and classification output.
Figure 3: T-branch three-path architecture detail. Path 1: Residual Conv-SE BiGRU (local and medium-range patterns). Path 2: Stride-Conv BiGRU (coarse-scale patterns). Path 3: Full-resolution pre-LayerNorm Transformer (global temporal context). All three paths merge onto a shared 8-step temporal grid before multi-head self-attention and mean-pooling to produce hT∈R512 .
Figure 4: CB-GAF mechanism detail for branch T as a representative example. Each branch projects to a common dimension df=128 , queries other two branches simultaneously via cross-attention, then applies a learned vector gate gT∈(0,1)128 to produce the gated residual fusion tfused . Identical structure applied in parallel for branches C and H.
Figure 5: Classification head with residual skip connection detail. The mean-pooled input xˉ∈R46 is projected to r∈R64 and concatenated with hfuse to form z∈R448 . A two-layer MLP with a residual skip Wskip:448→128 provides a direct gradient highway from the full input to the output layer, preventing gradient vanishing in the deep classification pathway.
Hyperparameter
Value
Rationale
Window W
32
Covers short attack bursts
Stride S
4
Coverage vs. compute
Embed dim de
32
Compact context representation
Conv channels
[64, 128, 128]
Progressive feature expansion
SE reduction ratio r
8
Channel recalibration efficiency
GRU hidden (Path 1)
128/dir, 2 layers
Matched to baselines
Table 6: Hyperparameter Settings
Model
F1
ROC-AUC
MCC
PR-AUC
Δ F1 (sig.)
TCH-Net (Ours)
0.8296±0.0028
0.9380±0.0025
0.6972±0.0056
0.8912±0.0031
—
BiLSTM-IDS [ 17 ]
0.7805±0.0010
0.8975±0.0001
0.5972±0.0030
0.8441±0.0018
+0.0491∗∗
BiGRU-IDS [ 12 ]
0.7805±0.0011
0.8962±0.0013
0.5987±0.0034
0.8438±0.0022
+0.0491∗∗
1D-CNN-IDS
0.7932±0.0076
0.9076±0.0062
0.6213±0.0153
0.8601±0.0091
+0.0364∗
Transformer-IDS [ 18 ]
0.7958±0.0030
0.9147±0.0012
0.6255±0.0067
0.8699±0.0041
+0.0338∗∗
MLP-IDS
0.7039±0.0008
0.8152±0.0005
0.4348±0.0018
0.7311±0.0012
+0.1257∗∗∗
Table 7: Main comparison results. Mean ± std across seeds. Best per metric in bold . Δ F1: absolute F1 improvement of TCH-Net over each baseline. Significance: ∗∗∗p<0.001 , ∗∗p<0.01 , ∗p<0.05 (one-sided paired Wilcoxon signed-rank test).
Figure 6: Radar chart comparing TCH-Net against the five strongest baselines across all four evaluation metrics. TCH-Net (filled area) consistently extends beyond all baselines on every axis.
Model
Params (M)
Latency (ms)
Throughput ( × 10 3 sps)
Mem (MB)
F1
TCH-Net (Ours) ⋆
2.692
6.43±0.18
20.5
10.27
0.8296
BiLSTM-IDS [ 17 ]
0.609
0.74±0.02
34.2
2.32
0.7805
BiGRU-IDS [ 12 ]
0.465
0.58±0.03
52.0
1.77
0.7805
1D-CNN-IDS
0.068
0.69±0.03
406.9
0.26
0.7932
Transformer-IDS [ 18 ]
0.618
1.22±0.03
36.8
2.36
0.7958
CNN-LSTM
0.142
0.88±0.05
273.5
0.54
0.7919
Table 8: Computational efficiency comparison. Latency: single-sample mean ± std over n=200 runs following 20 GPU warm-up passes, CUDA event timing, NVIDIA Tesla T4. Throughput: batch-512 inference, same device. Mem: runtime GPU memory footprint. F1: canonical mean from Table 7 . ⋆ = proposed model.
Variant
F1 ± std
AUC
MCC
Δ F1
TCH (Full)
0.8296±0.0028
0.9380
0.6972
—
T+C
0.7752±0.0012
0.8883
0.5880
−0.0544
T+H
0.7756±0.0014
0.8885
0.5870
−0.0540
C+H
0.7061±0.0003
0.8218
0.4402
−0.1235
T only
0.7753±0.0013
0.8882
0.5878
−0.0543
C only
0.6000±0.0000
0.4999
0.0000
−0.2296
Table 9: Branch ablation results (2 seeds each, seeds {42, 123}). Full model in bold.
Figure 7: Branch ablation F1, AUC, and MCC scores for all seven branch subsets. The dashed vertical line marks the full TCH-Net F1 ( =0.8296 ). All-branch fusion with CB-GAF provides the highest performance on every metric.
Variant
F1 ± std
AUC
MCC
Δ F1
Full TCH-Net
0.8296±0.0028
0.9380
0.6972
—
w/o CB-GAF
0.7759±0.0011
0.8898
0.5889
−0.0537
w/o MSTE (Three-Path Enc.)
0.7760±0.0019
0.8890
0.5903
−0.0536
w/o Aux. Loss
0.7755±0.0024
0.8893
0.5875
−0.0541
w/o All (v2)
0.7752±0.0022
0.8901
0.5857
−0.0544
Δ F1 values bound individual contributions (inclusive of proxy architectural gap); refer footnote † .
Table 10: Novelty component ablation (2 seeds, {42,123} ). Full model in bold.
Figure 8: Leave-one-dataset-out (LODO) F1 compared to in-distribution random-split F1 per held-out dataset. The dashed line marks the LODO mean ( =0.5577 ). The gap annotation on CICIDS-2017 shows the worst-case generalisation shortfall ( +0.6377 ), attributable to dataset dominance rather than feature coverage. † Bot-IoT test set n=38 (unreliable). ⋆ Supplementary dataset.
Model
LODO F1
± std
LODO AUC
Δ F1
BiLSTM-IDS [ 17 ]
0.3881
0.3062
0.4948
+ 0.1696
Transformer-IDS [ 18 ]
0.3910
0.2527
0.4831
+ 0.1667
BiGRU-IDS [ 12 ]
0.4449
0.2607
0.4948
+ 0.1128
1D-CNN-IDS
0.4592
0.2380
0.4105
+ 0.0985
CNN-LSTM
0.4654
0.2351
0.5500
+ 0.0923
Baseline mean
0.4297
—
0.4866
+ 0.1280
Table 14: Baseline LODO generalisation benchmark (2 seeds, same protocol as Table 13 ). Only the five deep learning baselines are included; classical ML baselines (RF, XGBoost) are omitted as their inferior in-distribution F1 makes their LODO performance a lower bound of limited analytical interest. Δ F1: TCH-Net LODO F1 (0.5577, from Table 13 ) minus baseline mean LODO F1. A positive Δ F1 indicates TCH-Net retains superior cross-dataset generalisation.
Cross-domain intrusion detection remains a critical challenge due to significant variability in network traffic characteristics and feature distributions across environments. This study evaluates the transferability of three widely used flow-based feature sets (Argus, Zeek and CICFlowMeter) across four widely used datasets representing heterogeneous IoT and Industrial IoT network conditions. Through extensive experiments, we evaluate in- and cross-domain performance across multiple classification models and analyze feature importance using SHapley Additive exPlanations (SHAP). Our results show that models trained on one domain suffer significant performance degradation when applied to a different target domain, reflecting the sensitivity of IoT intrusion detection systems to distribution shifts. Furthermore, the results evidence that the choice of classification algorithm and feature representations significantly impact transferability. Beyond reporting performance differences and thorough analysis of the transferability of features and feature spaces, we provide practical guidelines for feature engineering to improve robustness under domain variability. Our findings suggest that effective intrusion detection requires both high in-domain performance and resilience to cross-domain variability, achievable through careful feature space design, appropriate algorithm selection and adaptive strategies.
Alejandro Guerra-Manzanares, Jialin Huang
School of Computer Science, University of Nottingham, Ningbo, China · School of Computer Science, University of Nottingham, Nottingham, United Kingdom
Lightweight machine learning models are increasingly proposed for intrusion detection in Industrial Internet of Things (IIoT) networks due to their suitability for resource-constrained edge deployment. Most reported results evaluate these models only within their training network, leaving behavior on unseen networks unverified. This study trains four lightweight architectures on one IIoT dataset and evaluates them, without retraining, on two structurally distinct IIoT datasets using a feature representation restricted to attributes available across all three sources. Explainability analysis across two top-performing models shows both rely overwhelmingly on coarse port-category features; the most influential category occurs in source-domain attack traffic at 96 to 435 times the rate in the two target domains, indicating that coarsening port resolution relocates rather than removes a documented shortcut. Evaluation under naturally imbalanced class distributions reveals a further effect: the evaluation protocol used can reverse which target network appears to pose the greater generalization challenge. Adversarial robustness and recovery through limited target-domain exposure are also assessed; robustness to adversarial perturbation is unrelated to cross-network generalization, and recovery through adaptation varies considerably by architecture. These findings suggest deployment readiness should be assessed using cross-network evaluation under realistic class distributions, rather than within-domain accuracy alone.
MD Azizul Hakim, Md Shihab Uddin, Talha Ibne Anis
Department of Computer Science and Technology, Bangladesh Sweden Polytechnic Institute Kaptai, Rangamati, Chittagong 4530, Bangladesh · Department of Electrical Engineering, Bangladesh Sweden Polytechnic Institute Kaptai, Rangamati, Chittagong 4530, Bangladesh
With the rapid growth of interconnected devices in Industrial and Medical Internet of Things (IIoT and MIoT) ecosystems, ensuring timely and accurate detection of cyber threats has become a critical challenge. This study presents an advanced intrusion detection framework based on a hybrid Squeeze-and-Excitation Attention Vision Transformer-Bidirectional Long Short-Term Memory (SE ViT-BiLSTM) architecture. In this design, the traditional multi-head attention mechanism of the Vision Transformer is replaced with Squeeze-and-Excitation attention, and integrated with BiLSTM layers to enhance detection accuracy and computational efficiency. The proposed model was trained and evaluated on two real-world benchmark datasets; EdgeIIoT and CICIoMT2024; both before and after data balancing using the Synthetic Minority Over-sampling Technique (SMOTE) and RandomOverSampler. Experimental results demonstrate that the SE ViT-BiLSTM model outperforms existing approaches across multiple metrics. Before balancing, the model achieved accuracies of 99.11% (FPR: 0.0013%, latency: 0.00032 sec/inst) on EdgeIIoT and 96.10% (FPR: 0.0036%, latency: 0.00053 sec/inst) on CICIoMT2024. After balancing, performance further improved, reaching 99.33% accuracy with 0.00035 sec/inst latency on EdgeIIoT and 98.16% accuracy with 0.00014 sec/inst latency on CICIoMT2024.
Afrah Gueriani, Hamza Kheddar, Ahmed Cherif Mazari +2
LSEA Lab., Faculty of Technology, University of MEDEA, Medea 26000, Algeria · LSEA Lab, Faculty of Science, University of MEDEA, Medea 26000, Algeria · Department of Computer Engineering, Gazi University, Ankara, Turkey +1