Organizations: School of Cyber Science and Technology, Beihang University, Beijing, 100191, China · School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing, 100876, China · Beijing Electronic Science and Technology Institute, Beijing, 100070, China
To mitigate attention dilution in high-entropy TLS 1.3 flows, we propose BGA, a noise-immune neural distillation framework for encrypted threat intelligence.The methodology first employs Analysis of Variance (ANOVA) to decouple high-discriminatory control-plane features - specifically industrial setpoints - from stochastic cryptographic noise. To resolve the extreme class imbalance within a corpus of 86,878 flow records, a Wasserstein GAN with Gradient Penalty (WGAN-GP) module, enforcing the 1-Lipschitz constraint, is integrated to synthesize high-fidelity minority samples, elevating the detection recall of rare Malicious State Command Injections(MSCI) attacks by 43.2%. At its core, the BGA architecture integrates Bidirectional Long Short-Term Memory (BiLSTM) for temporal dependency extraction and an Adaptive Gated Multi-Head Attention mechanism. This gated unit functions as a neural filter to dynamically suppress encryption artifacts while amplifying malicious signatures. Extensive evaluations on CIC-IDS-2018 and Edge-IIoT benchmarks demonstrate a performance ceiling exceeding 95.2% across all key metrics. Furthermore, noise-injection stress tests confirm BGAs superior structural resilience with a 8.57% performance margin over vanilla Transformers, while its ultra-low inference latency of 0.2820 ms (estimated 1.6920 ms via theoretical scaling for ARM) indicates a high potential for real-time feasibility on heterogeneous industrial edge gateways, providing a promising architectural baseline for future hardware implementation.
Figures & tables
Feature
RNN / LSTM
Transformer
BiLSTM
BGA (Ours)
Temporal Context
Unidirectional
Global (Non-seq)
Bidirectional
Bidirectional + Gated
Representation
Raw Hidden State
Self-Attention
Concatenated
Distilled Signatures
Entropy Handling
Susceptible to noise
Attention Dilution
Feature Overlap
Neural Filter (Gated)
Data Imbalance
Majority-class bias
Majority-class bias
Standard training
WGAN-GP Optimized
Edge Suitability
High latency/Low acc.
Memory-intensive
Medium efficiency
Ultra-low (0.28 ms)
Table 1: In-depth comparison of BGA and experimental baseline architectures.
Rank
Feature Name
Physical Interpretation
F-Value
1
setpoint
Target Control Setpoint
350,466.82
2
control_scheme
Control Logic Configuration
261,368.60
3
resp_read_fun
Modbus Read Function Code
22,120.04
4
control_mode
System Operational Mode
7,359.59
5
command_address
Target Modbus Register Address
6,164.74
6
comm_read_function
Command Read Operation Type
5,682.57
Table 2: Feature importance ranking for the Edge-IIoT (Gas Pipeline) subset based on ANOVA F-values.
Figure 1: Statistical distribution (violin plot) of setpoint across traffic categories.
Attack Category
Original Samples
Augmented Samples
Increase Ratio
Normal Operation
48,910
48,910
1.0 ×
CMRI
12,371
24,455
1.9 ×
MSCI
623
24,455
39.2 ×
MPCI
6,128
24,455
4.0 ×
DoS
1,470
24,455
16.6 ×
Total
69,502
146,730
2.1 ×
Table 3: Class distribution of the Training Set (80% split) before and after WGAN-GP augmentation
Figure 2: WGAN-GP module architecture for data augmentation.
Figure 3: Proposed BGA model architecture integrating BiLSTM and Gated Attention.
Dimension
SMOTE + LSTM
BGA Model (Ours)
Feature Modeling
Simple attention/weighting
Gated Multi-Head Attention
Augmentation
Linear interpolation (SMOTE)
WGAN-GP (Non-linear)
Sequence Modeling
Unidirectional (limited context)
BiLSTM (Full context)
Robustness
Prone to majority-class bias
Superior Generalization
Table 4: Comparison between BGA and Traditional Models
Label
Attack Category
Raw Sample Count
Proportion (%)
0
Normal Operation
61,156
70.38
2
Complex Malicious Response Injection (CMRI)
15,466
17.80
4
Malicious Parameter Command Injection (MPCI)
7,637
8.79
6
Denial of Service (DoS)
1,837
2.11
3
Malicious State Command Injection (MSCI)
782
0.90
Total
-
86,878
100.0
Table 5: Raw Sample Distribution of the Edge-IIoT (Gas Pipeline) Subset.
Category
Attack Type
Sample Count
Proportion (%)
Benign
Normal Web/SSH/FTP encrypted traffic
611,560
88.08
DoS
DoS-Slowloris / DoS-GoldenEye
42,430
6.11
Infiltration
Exploiting vulnerable applications
16,190
2.33
Brute Force
FTP/SSH Brute Force attempts
15,280
2.20
Total
-
685,460
100.0
Table 6: Sample Distribution of the CIC-IDS-2018 Validation Subset.
Hyperparameter / Component
Value / Configuration
BiLSTM Layer
Hidden Units
64 (Bidirectional, 128 total)
Dropout Rate
0.2
Gated Multi-head Attention
Number of Heads ( h )
4
Latent Dimension per Head
32
Table 7: Detailed neural architecture and hyperparameter configurations for the BGA framework and WGAN-GP module.
Figure 4: Training loss dynamics (a) and multi-class confusion matrix on Edge-IIoT (b).
Dataset
Precision (%)
Recall (%)
F1-Score (%)
Accuracy (%)
CIC-IDS-2018
92.70
95.27
93.95
95.27
Edge-IIoT
90.83
95.26
92.99
95.26
Table 8: Support-based Weighted Average Performance on Different Datasets.
Figure 5: Weighted average performance metrics (Precision, Recall, F1-Score, Accuracy) on CIC-IDS-2018 and Edge-IIoT datasets.
Figure 6: Recall performance comparison across different sequence modeling architectures.
Model
Normal (%)
CMRI (%)
MSCI (%)
MPCI (%)
DoS (%)
BGA (Ours)
98.10
98.38
94.89
97.77
93.66
BiLSTM
94.87
93.97
97.01
97.46
92.33
LSTM
97.97
98.31
90.64
97.69
88.40
Transformer
98.03
98.53
92.07
97.78
59.89
RNN
97.95
98.59
94.74
97.82
86.85
Table 9: Detailed Per-class Recall Comparison on the Edge-IIoT Dataset (%)
Model Variant
Precision (%)
Recall (%)
F1-Score (%)
Accuracy (%)
Params (K)
Serial Latency (PC) (ms)
BaseLSTM
97.34 ± 0.03
97.27 ± 0.04
97.22 ± 0.05
97.27 ± 0.03
23.9
0.6130
BiLSTM
97.35 ± 0.06
97.29 ± 0.04
97.24 ± 0.07
97.29 ± 0.06
47.7
0.1625
BiLSTM+MHA
97.83 ± 0.12
97.77 ± 0.14
97.78 ± 0.13
97.77 ± 0.11
130.6
0.2930
BGA (Ours)
97.92 ± 0.02
97.86 ± 0.03
97.87 ± 0.03
97.86 ± 0.02
130.6
0.2820
Table 10: In-depth statistical ablation analysis across classification metrics and hardware efficiency (Mean ± SD over 5 runs).
Figure 7: Ablation study visualization focused on the high-precision regime [0.965, 0.985] across all evaluation metrics.
Figure 8: Multi-dimensional benchmarking of model variants: (a) classification fidelity, and (b) complexity-latency trade-off.
Model Architecture
Parameters (M)
Latency (ms)
Latency Ratio
TinyBERT (4-layers)
14.35 M
4.1946 ms
14.9 ×
BGA (Ours)
0.13 M
0.2820 ms
1.0 ×
Table 11: Quantitative feasibility comparison: BGA vs. Lightweight Transformer (TinyBERT).
Figure 9: Impact of WGAN-GP augmentation on classification fidelity: (a) original dataset vs. (b) augmented dataset.
Dataset
Normal (%)
CMRI (%)
MSCI (%)
MPCI (%)
DoS (%)
Original
98.11
98.38
51.67
97.46
92.33
Augmented
98.10 ± 0.08
98.38 ± 0.36
94.89 ± 1.13
97.77 ± 0.16
93.66 ± 3.38
Table 12: Impact of Data Augmentation on Class-wise Recall (%)
Environment
Core Configuration
Inference Latency
Throughput (PPS)
Research Workstation (PC)
Multi-core/GPU/2.4GHz+
0.2820 ms
3,546
Simulated Edge Gateway (ARM)
Single-core/1.2GHz
1.6920 ms
591.0
ICS Real-time Threshold
Deterministic Limit
¡ 10.0 ms
—
Table 13: Performance contrast for BGA between Research Workstation and Simulated Edge Gateway.
Figure 10: Gated attention heatmaps illustrating neural distillation patterns for Normal (top) and MPCI (bottom) traffic profiles.
Figure 11: BGA model hyperparameter sensitivity: F1-score variation across hidden dimensions (a) and attention heads (b).
Faculty of Computer Science & Information Technology Universiti Malaysia Sarawak (UNIMAS), Kota Samarahan, Malaysia · Founder & Lead Architect, Vixero Technology Enterprise, Kuching, Sarawak, Malaysia
School of Microelectronics and Communication Engineering, Chongqing University, Chongqing 400044, China · School of Data Science, Lingnan University, Hong Kong, China