README.md

---
license: apache-2.0
language:
- en
- de
---

# 🛡️ MLP Cybersecurity Classifier

This repository hosts a lightweight `scikit-learn`-based MLP classifier trained to distinguish cybersecurity-related content from other text, using sentence-transformer embeddings. It supports English and German input texts.

## 📊 Training Data

The model was trained on a multilingual dataset of cybersecurity and non-cybersecurity news articles. The dataset is publicly available on Zenodo:  
🔗 [https://zenodo.org/records/16417939](https://zenodo.org/records/16417939)

## 📦 Model Details

- **Architecture**: `MLPClassifier` with hidden layers `(128, 64)`
- **Embedding model**: [`intfloat/multilingual-e5-large`](https://huggingface.co/intfloat/multilingual-e5-large)
- **Input**: Cleaned article (removed stopwords) or report text
- **Output**: Binary label (e.g., `Cybersecurity`, `Not Cybersecurity`)
- **Languages**: English, German

## 🔧 Usage

```python
from sentence_transformers import SentenceTransformer
from huggingface_hub import hf_hub_download
import joblib

# 1. Load the embedding model
embedder = SentenceTransformer("intfloat/multilingual-e5-large")

# 2. Load the pretrained MLP classifier from Hugging Face Hub
model_path = hf_hub_download(repo_id="selfconstruct3d/cybersec_classifier", filename="cybersec_classifier.pkl")
model = joblib.load(model_path)

# 3. Example input texts (can be in English or German)
texts = [
    "A new ransomware attack has affected critical infrastructure in Germany.",
    "The local sports club hosted its annual summer festival this weekend."
]

# 4. Generate embeddings
embeddings = embedder.encode(texts, convert_to_numpy=True, show_progress_bar=False)

# 5. Predict cybersecurity relevance
predictions = model.predict(embeddings)

# 6. Output results
for text, label in zip(texts, predictions):
    print(f"Text: {text}\nPrediction: {label}\n")