moe-wordlist-dga-models

🛡️ Expert Models for Wordlist-Based DGA Detection

Systematic evaluation of seven expert models for detecting wordlist-based Domain Generation Algorithms (DGAs), identifying ModernBERT as the optimal expert achieving 86.7% F1-score on known families and 80.9% on unseen variants.

---

📋 Overview

This repository contains the complete implementation of expert model evaluation for wordlist-based DGA detection, as described in our research paper (currently under review). Wordlist-based DGAs generate linguistically coherent domains that evade traditional detection methods, making them particularly challenging for cybersecurity systems.

🎯 Key Findings

Key Improvement: Specialist training provides +9.4% F1-score improvement over generalist approaches on known families and +30.2% on unseen families.

---

🚀 Quick Start

Option 1: Use Pre-trained Model (Recommended)

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load the optimal expert model
model_name = "Reynier/moe-wordlist-dga-models"
model_path = f"{model_name}/models/modernbert-wordlist-expert"

tokenizer = AutoTokenizer.from_pretrained("answerdotai/ModernBERT-base")
model = AutoModelForSequenceClassification.from_pretrained(
    model_name,
    subfolder="models/modernbert-wordlist-expert"
)

# Classify a domain
domain = "secure-banking-portal.com"
inputs = tokenizer(domain, return_tensors="pt", truncation=True, max_length=128)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=1)

print(f"Benign: {probs[0][0]:.4f}")
print(f"DGA:    {probs[0][1]:.4f}")

Option 2: Clone and Explore

# Clone the repository
git clone https://huggingface.co/Reynier/moe-wordlist-dga-models
cd moe-wordlist-dga-models

# Explore available models
ls models/

# Load datasets
python3 << EOF
import csv
with open('datasets/train_wl.csv', 'r') as f:
    reader = csv.DictReader(f)
    for i, row in enumerate(reader):
        if i < 5:
            print(row)
EOF

---

📦 Repository Contents

🤖 Models (7 Expert Candidates)

All models are located in models/ directory:

1. modernbert-wordlist-expert/ ⭐ (OPTIMAL)

   • Base: answerdotai/ModernBERT-base
   • F1-score: 86.7% (known), 80.9% (unknown)
   • Inference: 26ms on Tesla T4
   • Size: 575 MB

2. modernbert-generalist-54f/ (Baseline)

   • Trained on 54 diverse DGA families
   • F1-score: 79.2% (known), 62.1% (unknown)
   • Demonstrates specialist advantage

3. gemma-3-4b-lora/

   • LoRA adapters for google/gemma-3-4b-it
   • Exceptional precision (95.4%), lower recall (66.5%)
   • Size: 95 MB (adapters only)

4. llama-3.2-3b-lora/

   • LoRA adapters for meta-llama/Llama-3.2-3B
   • Balanced performance, slow inference
   • Size: 110 MB (adapters only)

5. dombert-url/

   • Domain-specialized BERT variant
   • Strong generalization (84.6% on unknown)
   • Size: 1.4 MB (LoRA adapters)

6. cnn-wordlist/

   • Convolutional neural network
   • Fastest inference (15ms), moderate accuracy
   • Size: 76 KB

7. fanci/

   • Random Forest with engineered features
   • Traditional ML baseline
   • Size: 794 MB (includes dictionaries)

8. labin/

   • Hybrid linguistic-attention model
   • Keras implementation
   • Size: 8.1 MB

📊 Datasets

All datasets are in datasets/ directory:

Training Datasets

1. train_wl.csv (160,000 samples)

   • Purpose: Train expert models (specialist approach)
   • DGA Families (8): charbot, deception, gozi, manuelita, matsnu, nymaim, rovnix, suppobox (10K each)
   • Benign: 80,000 domains from Tranco top sites
   • Distribution: Balanced (50% DGA, 50% benign)
   • Format: domain,family,label

2. train_1M.csv (1,080,000 samples)

   • Purpose: Train generalist model (baseline comparison)
   • DGA Families: 54 diverse families (wordlist + algorithmic)
   • Distribution: Diverse multi-family dataset

Test Datasets

1. test-known/ (8 families)

   • Purpose: Evaluate performance on training families
   • Total samples: 723,847
   • Families:
     • charbot: 11,001 samples
     • deception: 30,001 samples
     • gozi: 50,212 samples
     • manuelita: 20,001 samples
     • matsnu: 116,480 samples
     • nymaim: 217,773 samples
     • rovnix: 120,351 samples
     • suppobox: 158,028 samples
   • Format: Compressed .gz files (one per family)

2. test-generalization/ (3 families)

   • Purpose: Test generalization to unseen wordlist-based DGAs
   • Total samples: 13,562
   • Families:
     • bigviktor: 2,001 samples
     • ngioweb: 2,001 samples
     • pizd: 9,560 samples
   • Format: Compressed .gz files

📓 Notebooks

Training and evaluation notebooks in notebooks/:

• ModernBERT_base_DGA_Word.ipynb - Train ModernBERT expert (8 families)
• ModernBERT_base_DGA_54F.ipynb - Train ModernBERT generalist (54 families)
• Train_Gemma3_4B_DGA_WordList.ipynb - Fine-tune Gemma with LoRA
• Train_llama3B_DGA_WordList.ipynb - Fine-tune LLaMA with LoRA
• Test_Gemma3_4B_DGA_Last.ipynb - Evaluate Gemma model
• Test__llama3B_DGA.ipynb - Evaluate LLaMA model
• DomUrlBert.ipynb - Train DomBertUrl model
• CNN_Patron_WL.ipynb - Train CNN model
• FANCI.ipynb - Train FANCI Random Forest
• Labin_wl.ipynb - Train LABin model

---

🔬 Research Methodology

Two-Phase Evaluation Protocol

1. Phase 1: Known Families Performance

   • Evaluate on 8 training families
   • 30 test batches × 100 domains per family
   • Measures detection accuracy on familiar variants

2. Phase 2: Generalization Capability

   • Evaluate on 3 unseen wordlist-based families
   • Tests robustness against novel DGA variants
   • Critical for real-world deployment

Evaluation Metrics

• Precision: Accuracy of DGA predictions
• Recall: Coverage of actual DGAs
• F1-Score: Harmonic mean (primary metric)
• False Positive Rate (FPR): Benign misclassification rate
• Inference Time: Real-world performance (Tesla T4 GPU)

DGA Families

Training Families (8 wordlist-based):

• charbot, deception, gozi, manuelita, matsnu, nymaim, rovnix, suppobox

Generalization Test (3 wordlist-based):

• bigviktor, ngioweb, pizd

---

🛠️ Installation & Requirements

Basic Requirements

pip install torch transformers safetensors

For LLM Models (Gemma/LLaMA)

pip install peft accelerate bitsandbytes

For Traditional Models (FANCI)

pip install scikit-learn joblib

For LABin Model

pip install tensorflow keras

GPU Recommendations

• Optimal: NVIDIA Tesla T4 or better
• Minimum: 8GB VRAM for ModernBERT
• LLMs: 16GB+ VRAM (or use 8-bit quantization)

---

📊 Reproducibility

All experiments are fully reproducible:

1. Download datasets from datasets/ folder
2. Run training notebooks from notebooks/
3. Load pre-trained models from models/
4. Verify reported metrics using test sets

Expected Results (±std)

---

📖 Citation

This work is currently under review. Preliminary citation:

@article{leyva2025expert,
  title={Expert Selection for Wordlist-Based DGA Detection: A Systematic Evaluation},
  author={Leyva La O, Reynier and Catania, Carlos A. and Gonzalez, Rodrigo},
  journal={Under Review},
  year={2025}
}

---

🔗 Links

• GitHub Repository: MoE-word-list-dga-detection
• Paper: Under Review
• Contact: [email protected]

---

📄 License

This project is licensed under the MIT License - see LICENSE file for details.

---

🙏 Acknowledgments

• Datasets: DGArchive, 360 Netlab, UMUDga, Tranco
• Base Models: ModernBERT (Answer.AI), Gemma (Google), LLaMA (Meta)
• Infrastructure: CONICET Argentina

---

🔍 Quick Navigation

• Models Directory
• Datasets Directory
• Training Notebooks
• Results & Metrics

---

Last Updated: October 2025