---
license: apache-2.0
language:
  - en
library_name: pytorch
pipeline_tag: text-classification
datasets:
  - Ammar-ss/BRIDGE
tags:
  - network-security
  - iot-security
  - intrusion-detection
  - anomaly-detection
  - botnet-detection
  - cybersecurity
  - gated-attention
  - multi-branch
metrics:
  - f1
  - roc_auc
  - matthews_correlation
  - accuracy
arxiv: 2604.11324
model-index:
  - name: TCH-Net
    results:
      - task:
          type: text-classification
          name: IoT Botnet Detection
        dataset:
          name: BRIDGE
          type: Ammar-ss/BRIDGE
        metrics:
          - type: f1
            value: 0.8296
            name: F1 (macro, 5 seeds)
          - type: roc_auc
            value: 0.9380
            name: ROC-AUC (5 seeds)
          - type: matthews_correlation
            value: 0.6972
            name: MCC (5 seeds)
          - type: f1
            value: 0.5577
            name: LODO F1 (leave-one-dataset-out)
---

# TCH-Net: Multi-Branch IoT Botnet Detection on BRIDGE

**Paper:** [BRIDGE and TCH-Net: Heterogeneous Benchmark and Multi-Branch Baseline for Cross-Domain IoT Botnet Detection](https://arxiv.org/abs/2604.11324)  
**Authors:** Ammar Bhilwarawala, Likhamba Rongmei, Harsh Sharma, Arya Jena, Kaushal Singh, Jayashree Piri, Raghunath Dey — KIIT University  
**Submitted to:** Journal of Network and Computer Applications  
**Dataset:** [Ammar-ss/BRIDGE](https://huggingface.co/datasets/Ammar-ss/BRIDGE)  
**Code:** [github.com/Ammar-ss/TCH-Net](https://github.com/Ammar-ss/TCH-Net)

---

## What is this?

The IoT botnet detection field has a quiet problem. Almost every published system gets trained on one dataset, reports numbers in the high 90s, and calls it done. The trouble is those numbers don't travel. A model tuned to CICIDS-2017 will see completely different traffic statistics when you point it at Bot-IoT or N-BaIoT — different capture tools, different devices, different attack toolkits. The benchmark looked easy because it was a closed world.

TCH-Net is a multi-branch neural network built to handle this more honestly. It's trained and evaluated on **BRIDGE**, a unified benchmark that maps five structurally distinct public datasets into a shared 46-feature space. The goal was to build something that could survive being tested on genuinely heterogeneous data — and then to measure exactly how hard that actually is.

The architecture has three parallel branches. The **Temporal branch (T)** runs three paths simultaneously: a residual depthwise-separable convolutional BiGRU for local and medium-range patterns, a stride-downsampled BiGRU for coarser dynamics, and a full-resolution pre-LayerNorm Transformer covering all 32 timesteps for global context. Different botnet attack categories manifest at different temporal scales — DDoS flooding shows up in burst-level signatures, C&C beaconing in medium-scale periodic patterns, scan-then-exploit sequences in global ordering — and a single-resolution encoder has to trade one against the others. Three paths running in parallel resolves that.

The **Statistical branch (H)** mean-pools the window and runs it through an MLP. It captures distributional structure that doesn't depend on ordering at all, which is orthogonal to what the T-branch does. The **Contextual branch (C)** encodes the source dataset and device category as learned embeddings. On its own it's nearly random (AUC ≈ 0.50) — it doesn't predict attack labels independently. What it does is condition the fusion mechanism on where the input came from.

All three branches get fused through **CB-GAF** (Cross-Branch Gated Attention Fusion). Each branch queries the other two simultaneously via cross-attention, then a learned sigmoid vector gate — 128-dimensional, not a scalar — controls feature-wise how much cross-branch information gets absorbed. That vector gating is important in the heterogeneous setting. For a dataset like N-BaIoT where 85% of canonical features are zero-filled, the gate can learn to downweight the largely empty H-branch at specific dimensions rather than hard-coding that decision or just averaging in noise across the board.

**2,692,696 parameters.** Inference latency 6.43ms on Tesla T4. Fits on NVIDIA Jetson with room to spare.

---

## Intended Use

### In-Scope
- IoT botnet detection research on network flow data
- Cross-dataset generalisation benchmarking in heterogeneous IDS settings
- Ablation or architecture comparison studies using the BRIDGE benchmark

### Out-of-Scope
- **Production deployment without retraining:** The LODO gap (0.2719 F1) indicates the model does not reliably generalise to unseen dataset distributions without adaptation. Do not deploy this in a live network without fine-tuning on in-distribution data.
- **Non-IoT or enterprise network traffic:** BRIDGE covers IoT-specific datasets. Behaviour on corporate LAN/WAN traffic is not evaluated.
- **Real-time per-packet classification:** TCH-Net operates on flow-level feature windows of 32 timesteps. It requires completed or windowed flows, not individual packets.
- **Unknown context at scale:** The contextual branch requires dataset source and device category IDs. If these are unavailable, pass `ctx = torch.zeros(B, 2, dtype=torch.long)` — performance will degrade modestly but gracefully.

### Limitations
- LODO F1 = 0.5577. The model does not generalise well to unseen dataset distributions. This is the best LODO result across all evaluated architectures (+0.09–0.17 above baselines), but the gap is real and quantified.
- N-BaIoT achieves high F1 (0.9854) largely because Mirai/BASHLITE signatures are statistically distinct in only 7 of 46 features. This is not a BRIDGE-wide pattern.
- Edge-IIoTset is the hardest case (F1 = 0.6755) due to IIoT packet-level traffic structures differing from the flow-level distributions dominating training.
- 85% zero-fill on N-BaIoT canonical features is a BRIDGE artefact — the canonical feature space was built around CICIDS-style flow features, which do not map cleanly to all source datasets.
- Not validated on live traffic captures or real-world deployment scenarios.

---

## Results (5 seeds: 42, 123, 456, 789, 2024)

| Metric | TCH-Net | Best Baseline (Transformer-IDS) |
|---|---|---|
| **F1** | **0.8296 ± 0.0028** | 0.7958 ± 0.0030 |
| **ROC-AUC** | **0.9380 ± 0.0025** | 0.9147 ± 0.0012 |
| **MCC** | **0.6972 ± 0.0056** | 0.6255 ± 0.0067 |
| **PR-AUC** | **0.8912 ± 0.0031** | 0.8699 ± 0.0041 |

TCH-Net outperforms all 12 baselines on all four metrics. All differences are statistically significant (p < 0.05, one-sided paired Wilcoxon signed-rank test).

### Full Comparison Table

| Model | F1 | ROC-AUC | MCC | ΔF1 |
|---|---|---|---|---|
| **TCH-Net (Ours)** | **0.8296 ± 0.0028** | **0.9380** | **0.6972** | — |
| Transformer-IDS | 0.7958 ± 0.0030 | 0.9147 | 0.6255 | +0.0338** |
| 1D-CNN-IDS | 0.7932 ± 0.0076 | 0.9076 | 0.6213 | +0.0364* |
| CNN-LSTM | 0.7919 ± 0.0137 | 0.9056 | 0.6208 | +0.0377* |
| BiLSTM-IDS | 0.7805 ± 0.0010 | 0.8975 | 0.5972 | +0.0491** |
| BiGRU-IDS | 0.7805 ± 0.0011 | 0.8962 | 0.5987 | +0.0491** |
| DeepDefense | 0.7627 ± 0.0011 | 0.8776 | 0.5638 | +0.0669*** |
| XGBoost | 0.7265 ± 0.0014 | 0.8704 | 0.5542 | +0.1031*** |
| GraphSAGE-Approx | 0.7097 ± 0.0004 | 0.8259 | 0.4465 | +0.1199*** |
| Kitsune-AE | 0.7045 ± 0.0007 | 0.8200 | 0.4362 | +0.1251*** |
| MLP-IDS | 0.7039 ± 0.0008 | 0.8152 | 0.4348 | +0.1257*** |
| IoT-DNN | 0.7009 ± 0.0002 | 0.8146 | 0.4278 | +0.1287*** |
| Random Forest | 0.4323 ± 0.0082 | 0.8005 | 0.3557 | +0.3973*** |

### Per-Dataset Performance

| Dataset | Coverage | DetRate | False Alarm | F1 |
|---|---|---|---|---|
| CICIDS-2017 | 93% | 0.9433 | 0.0309 | 0.9505 |
| CIC-IoT-2023 | 87% | 0.8827 | 0.0257 | 0.9211 |
| N-BaIoT | 15% | 0.9982 | 0.0206 | **0.9854** |
| Edge-IIoTset | 22% | 0.6844 | 0.2589 | 0.6755 |

N-BaIoT achieves the highest F1 despite 85% of features being zero-filled. Mirai and BASHLITE botnet traffic is statistically distinctive enough in just 7 features that the separation is stark. Edge-IIoTset is the hardest case — IIoT packet-level traffic structures differently from the flow-level distributions that dominate training.

### Leave-One-Dataset-Out (LODO) Generalisation

The honest number. Train on four datasets, test on the fifth, repeated five times.

| Held-Out | LODO F1 | LODO AUC |
|---|---|---|
| CICIDS-2017 | 0.3128 ± 0.232 | 0.0509 |
| CIC-IoT-2023 | 0.6013 ± 0.000 | 0.1440 |
| Bot-IoT | 0.5934 ± 0.011 | 0.5693 |
| Edge-IIoTset | 0.6791 ± 0.008 | 0.6841 |
| N-BaIoT | 0.6021 ± 0.000 | 0.8171 |
| **MEAN** | **0.5577** | **0.4531** |

Generalisation gap: random-split F1 (0.8296) − LODO mean (0.5577) = **+0.2719**.

This gap is not a TCH-Net problem. All five deep learning baselines scored between 0.39–0.47 LODO F1 under the same protocol. TCH-Net's 0.5577 is the highest LODO score across all evaluated architectures — +0.09 to +0.17 above baselines. The gap is a measurement of how hard the cross-dataset problem actually is. The BRIDGE LODO mean of 0.5577 is the first formally quantified community generalisation baseline in heterogeneous IoT intrusion detection.

### Temporal Split Check

| Split | F1 | AUC | MCC |
|---|---|---|---|
| Random (5 seeds) | 0.8296 | 0.9380 | 0.6972 |
| Temporal (1 seed) | 0.8203 | 0.9261 | 0.6831 |
| Δ | −0.0093 | −0.0119 | −0.0141 |

The small drop under temporal splitting confirms TCH-Net's performance is not driven by temporal leakage.

---

## Architecture

```
Input: (B, 32, 46)  — batch × window × canonical features

Shared Feature Projection (residual):
  Linear(46→92) → LayerNorm → GELU → Dropout(δ/2)
  → Linear(92→46) → LayerNorm        X̃ = X + f_proj(X)

T-Branch — three parallel paths, merged to 512d:
  Path 1: ResConvSE×3 + MaxPool → BiGRU(128/dir, 2L)              → 8×256
  Path 2: StrideConv(s=2,64ch) → BiGRU(64/dir, 1L) → AvgPool(8)  → 8×128
  Path 3: Linear(46→128) + LearnablePE → TransEnc(Pre-LN,2L,8H)
           → strip CLS → AvgPool(8)                                → 8×128
  Merge:  concat → LayerNorm(512) → MHA(8 heads) → mean pool       → 512d

H-Branch:   mean(X̃, dim=time) → MLP(46→128→64, BN+GELU+Dropout)  → 64d

C-Branch:   Embed_ds(5,32)[c_ds] ‖ Embed_dev(6,32)[c_dev]         → 64d

CB-GAF (Cross-Branch Gated Attention Fusion):
  Project each branch to d_f=128
  Each branch queries both others simultaneously via cross-attention
  Per-branch vector gate g^i ∈ (0,1)^128  (feature-wise, not scalar)
  x_fused = g^i ⊙ x_self + (1−g^i) ⊙ x_cross
  concat(T_fused, C_fused, H_fused) → LayerNorm                   → 384d

Classifier (residual head):
  raw_proj(mean(X̃)) → 64d
  concat(384d fused, 64d raw) → z ∈ 448d
  Linear(448→256) → BN+GELU+Dropout
  Linear(256→128) + Wskip(448→128)  ← residual skip
  Linear(128→2) → softmax

Aux Decoder (training only, λ=0.05):
  MLP(384→64→46) — prevents information collapse in CB-GAF
```

---

## Ablation Summary

### Branch Ablation (2 seeds, CB-GAF replaced with concat)

| Variant | F1 | ΔF1 |
|---|---|---|
| **T+C+H (Full)** | **0.8296** | — |
| T+H | 0.7756 | −0.0540 |
| T+C | 0.7752 | −0.0544 |
| T only | 0.7753 | −0.0543 |
| H only | 0.7054 | −0.1242 |
| C+H | 0.7061 | −0.1235 |
| C only | 0.6000 | −0.2296 |

The C-branch alone gets near-random performance (AUC ≈ 0.50) — confirming its role is fusion conditioning, not independent prediction.

### Novelty Component Ablation (2 seeds)

| Variant | F1 | ΔF1 |
|---|---|---|
| **Full TCH-Net** | **0.8296** | — |
| w/o CB-GAF | 0.7759 | −0.0537 |
| w/o MSTE (Three-Path) | 0.7760 | −0.0536 |
| w/o Aux Loss | 0.7755 | −0.0541 |
| w/o All (v2 baseline) | 0.7752 | −0.0544 |

Removing any single novel component costs ~0.054 F1.

---

## Computational Profile (NVIDIA Tesla T4)

| Model | Params | Latency | Throughput | Memory | F1 |
|---|---|---|---|---|---|
| **TCH-Net** | 2.692M | 6.43±0.18ms | 20.5k sps | 10.27MB | **0.8296** |
| BiLSTM-IDS | 0.609M | 0.74±0.02ms | 34.2k sps | 2.32MB | 0.7805 |
| Transformer-IDS | 0.618M | 1.22±0.03ms | 36.8k sps | 2.36MB | 0.7958 |
| 1D-CNN-IDS | 0.068M | 0.69±0.03ms | 406.9k sps | 0.26MB | 0.7932 |

The 6.43ms latency supports 20,000+ detections/second under batch processing. The 10.27MB footprint is deployable on NVIDIA Jetson hardware. For microcontroller-class endpoints (Cortex-M, ESP32), quantisation or knowledge distillation would be needed.

---

## Installation

```bash
pip install torch scikit-learn numpy huggingface_hub
```

Tested with Python 3.9+. No specific version pinning required beyond a modern PyTorch (≥2.0 recommended for `torch.compile` compatibility).

---

## Loading the Model

```python
import torch
import pickle
import numpy as np
import torch.nn.functional as F
from huggingface_hub import hf_hub_download

# ── Download files ──────────────────────────────────────────────────────────
ckpt_path   = hf_hub_download("Ammar-ss/BRIDGE_and_TCH-Net", "tch_net_best.pth")
scaler_path = hf_hub_download("Ammar-ss/BRIDGE_and_TCH-Net", "scaler.pkl")

# ── Load scaler ──────────────────────────────────────────────────────────────
with open(scaler_path, "rb") as f:
    scaler = pickle.load(f)

# ── Load checkpoint ──────────────────────────────────────────────────────────
ckpt   = torch.load(ckpt_path, map_location="cpu")
config = ckpt["config"]

# ── Define TCHNet ─────────────────────────────────────────────────────────────
# Full class definition is in bridge-and-tch-net.ipynb and the GitHub repo.
# Paste or import TCHNet before instantiating:
#   from tch_net import TCHNet   (if using the GitHub repo)
#   OR copy the class from the notebook.

model = TCHNet(
    nf=config["n_features"],   # 46
    ws=config["window_size"],  # 32
    nc=config["n_classes"],    # 2
)
model.load_state_dict(ckpt["state_dict"])
model.eval()

# ── Preprocess ───────────────────────────────────────────────────────────────
# X_raw: np.ndarray of shape (N, 46) — raw canonical flow features
X_scaled = np.clip(scaler.transform(X_raw), -10, 10).astype(np.float32)

# ── Inference ────────────────────────────────────────────────────────────────
# x:   FloatTensor (B, 32, 46) — windowed, scaled flow features
# ctx: LongTensor  (B, 2)      — [dataset_source_id, device_category_id]
#
#   dataset_source_id:  0=CICIDS-2017  1=CIC-IoT-2023  2=Bot-IoT
#                       3=Edge-IIoTset  4=N-BaIoT
#   device_category_id: 0=sensor  1=camera  2=appliance  3=IIoT
#                       4=server  5=unknown
#
#   If context is unknown: ctx = torch.zeros(B, 2, dtype=torch.long)
#   The C-branch has no independent predictive power — unknown context
#   degrades gracefully, it does not break inference.

with torch.no_grad():
    logits, _ = model(x, ctx)
    probs = F.softmax(logits, dim=-1)
    preds = logits.argmax(dim=-1)  # 0 = benign, 1 = attack
```

---

## Files

| File | Description |
|---|---|
| `tch_net_best.pth` | Best checkpoint (highest F1 across all 5 seeds) |
| `tch_net_seed_42.pth` | Per-seed checkpoint, seed 42 |
| `tch_net_seed_123.pth` | Per-seed checkpoint, seed 123 |
| `tch_net_seed_456.pth` | Per-seed checkpoint, seed 456 |
| `tch_net_seed_789.pth` | Per-seed checkpoint, seed 789 |
| `tch_net_seed_2024.pth` | Per-seed checkpoint, seed 2024 |
| `scaler.pkl` | RobustScaler (q5–q95) fitted on BRIDGE training split — required for inference |
| `manifest.json` | Config, per-seed metrics, feature names |
| `BRIDGE and TCH-Net (FULL PAPER).ipynb` | Complete experimental notebook (all 12 baselines, branch ablation, novelty ablation, LODO, temporal split, adversarial robustness, HP sensitivity) |
| `bridge-and-tch-net.ipynb` | Clean training-only notebook (TCH-Net, 5 seeds, saves checkpoints) |

---

## Training Hyperparameters

| Parameter | Value |
|---|---|
| Optimizer | AdamW |
| Learning rate | 5×10⁻⁴ |
| Weight decay | 5×10⁻⁵ |
| Scheduler | Cosine annealing, 2-epoch warmup |
| Loss | Focal (γ=2.0, α-weighted, ε=0.05) + Aux (λ=0.05) |
| Batch size | 512 |
| Max epochs / patience | 30 / 5 |
| Sequence length | W=32, stride S=4 |
| Dropout | 0.15 |
| Input augmentation | Gaussian noise (σ=0.01, p=0.30, train only) |
| AMP | fp16 on CUDA |

---

## Citation

```bibtex
@article{bhilwarawala2026bridge,
  title   = {{BRIDGE} and {TCH-Net}: Heterogeneous Benchmark and Multi-Branch
             Baseline for Cross-Domain {IoT} Botnet Detection},
  author  = {Bhilwarawala, Ammar and Rongmei, Likhamba and Sharma, Harsh
             and Jena, Arya and Singh, Kaushal and Piri, Jayashree and Dey, Raghunath},
  journal = {arXiv preprint arXiv:2604.11324},
  year    = {2026}
}
```

---

## Model Card Authors

Ammar Bhilwarawala, KIIT University.  
For questions or issues, open a discussion on this repository.