Table 2 Configuration of the hybrid transformer–CNN intrusion detection model.
Component | Layer type | Configuration details | Purpose |
|---|---|---|---|
Input | Traffic Window | Fixed-length network traffic window | Temporal segmentation of MANET traffic |
Transformer Branch | Encoder Layers | 2 Transformer encoder layers | Capture long-range temporal dependencies |
Self-Attention | 4 attention heads | Learn global traffic correlations | |
Feed-Forward | Dense projection | Temporal feature refinement | |
CNN Branch | Convolution Layer 1 | Kernel size: 3 × 3 | Local spatial feature extraction |
Convolution Layer 2 | Kernel size: 5 × 5 | Multi-scale anomaly detection | |
Convolution Layer 3 | Kernel size: 3 × 3 | Fine-grained spatial representation | |
Pooling | Max pooling | Dimensionality reduction | |
Feature Fusion | Concatenation | F = concat(G, S) | Combine temporal and spatial features |
Classification | Fully Connected Layers | 2 FC layers | Feature abstraction |
Output Layer | Softmax | Multi-class classification | Attack type prediction |
Response Mechanism | Isolation Module | Route exclusion of malicious nodes | Network protection |
