Paste any headline or article. Get an instant REAL / FAKE verdict with confidence score and linguistic evidence.
Misinformation spreads 6x faster than the truth. FakeShield uses NLP and machine learning to classify news articles in milliseconds — with a confidence score, fake/real probability breakdown, and the exact words that triggered the verdict.
|
🧠 Multi-Model ML
|
🔍 Explainable AI
|
|
📊 Training Visualisations
|
🌐 REST API
|
git clone https://github.com/aasimansari1/fake-news-detector.git
cd fake-news-detector
pip install -r requirements.txt
# Generate sample dataset + download NLTK data
python dataset/create_dataset.py
# Train all 4 models (best one saved automatically)
python train_model.py
# Launch the web app
python app.py
# → http://localhost:5000Want better accuracy? Replace
dataset/sample_data.csvwith the Kaggle Fake News dataset (~44K articles). Any CSV withtextandlabelcolumns (REAL/FAKE or 0/1) works.
📰 Article / Headline
│
▼
┌───────────────────┐
│ NLP Pipeline │
│ lowercase → │
│ punctuation → │
│ tokenise → │
│ stopwords → │
│ lemmatise │
└────────┬──────────┘
│
▼
┌───────────────────┐
│ TF-IDF (bigrams) │
│ Vectoriser │
└────────┬──────────┘
│
▼
┌───────────────────┐
│ Best ML Model │ ← Passive Aggressive (~91%)
└────────┬──────────┘
│
▼
✅ REAL (confidence: 94%)
❌ FAKE (confidence: 82%)
+ key words + probabilities
POST /predict
Content-Type: application/json
{"text": "Scientists discover that vaccines cause autism, government hiding truth"}{
"prediction": "FAKE",
"confidence": 89.4,
"fake_probability": 89.4,
"real_probability": 10.6,
"key_words": [
{"word": "hiding", "score": 0.21, "influence": "fake"},
{"word": "truth", "score": 0.18, "influence": "fake"}
],
"model_name": "Passive Aggressive",
"model_accuracy": 0.907
}| Model | Accuracy | Notes |
|---|---|---|
| Passive Aggressive | ~91% | Best overall — auto-selected |
| Naive Bayes | ~88% | Fastest inference |
| Logistic Regression | ~86% | Most interpretable |
| Random Forest | ~84% | Most robust to noise |
Accuracy on sample dataset. Use a larger real-world dataset for production-grade results.
fake-news-detector/
├── app.py # Flask web app + /predict API
├── train_model.py # Training pipeline, model comparison
├── requirements.txt
│
├── src/ # Reusable data science module
│ ├── preprocessing.py # Text cleaning, label normalisation, dataset loader
│ ├── features.py # TF-IDF vectoriser builder
│ └── models.py # Train all classifiers, pick best, ROC data
│
├── notebooks/ # Jupyter data science pipeline
│ ├── 01_EDA.ipynb # Exploratory data analysis
│ ├── 02_Feature_Engineering.ipynb # Preprocessing + TF-IDF
│ ├── 03_Model_Training.ipynb # Train & compare 4 models
│ └── 04_Model_Evaluation.ipynb # Confusion matrix, ROC, feature importance, CV
│
├── data/
│ ├── raw/ # Source datasets (CSV)
│ └── processed/ # Train/test splits + vectoriser (generated)
│
├── models/ # Saved model + vectoriser (generated)
├── reports/figures/ # All visualisation outputs (generated)
│
├── dataset/
│ └── create_dataset.py # Sample dataset generator
├── static/
│ ├── css/style.css
│ ├── js/main.js
│ └── images/
└── templates/
└── index.html
pip install -r requirements.txt
jupyter notebookRun in order: 01_EDA → 02_Feature_Engineering → 03_Model_Training → 04_Model_Evaluation
| Layer | Technology |
|---|---|
| Backend | Python, Flask, Gunicorn |
| ML / NLP | scikit-learn, NLTK, pandas, numpy |
| Visualisation | matplotlib, seaborn |
| Frontend | Vanilla HTML/CSS/JS, dark mode |
| Server | Nginx + systemd |
Ideas for contributions:
- 🌐 Browser extension that checks articles in-page
- 🤗 Fine-tune a BERT/RoBERTa model for better accuracy
- 📱 Mobile-friendly UI improvements
- 🗃️ Support for more languages
git checkout -b feature/your-feature
git commit -m 'Add your feature'
git push origin feature/your-feature
# Open a Pull RequestMIT © Mohd Aasim Ansari
Fighting misinformation one article at a time. If this helped, please ⭐ star the repo!