Fake News Radar

Inspiration

In today’s digital era, misinformation spreads faster than truth. Fake news has led to misinformation, panic, and loss of trust in online platforms. I wanted to build an AI-powered solution that helps users quickly verify the authenticity of news before believing or sharing it.

What I Learned

Fine-tuning NLP models like BERT/DistilBERT for text classification.
Building a real dataset pipeline using public fake news datasets (e.g., LIAR, FakeNewsNet).
Implementing explainable AI by highlighting suspicious keywords.
Deploying machine learning models as APIs and integrating them with a clean, modern frontend.

How I Built the Project

Data Collection & Preprocessing

Collected datasets containing real and fake news headlines.
Performed tokenization, stopword removal, and text vectorization using BERT embeddings.

Model Training

Fine-tuned DistilBERT for binary classification: Real vs Fake.
Achieved over 90% F1-score using PyTorch and HuggingFace Transformers.

Backend

Developed a Flask API exposing a /predict endpoint that returns classification results and confidence scores.

Frontend

Built a responsive web UI using HTML, TailwindCSS, and JavaScript.
Implemented real-time predictions with AJAX requests to the Flask backend.

Deployment

Hosted the model using Hugging Face Spaces (Gradio) for interactive demos.
Deployed frontend on Vercel.

Challenges I Faced

Data Imbalance: Many datasets had more real news than fake news. I used data augmentation and resampling techniques to balance them.
Subtle Language Patterns: Fake news often looks authentic. I overcame this by fine-tuning contextual embeddings.
Model Size & Latency: BERT is large; I optimized using DistilBERT for faster inference.
Explainability: Building a heatmap of important words required custom visualization tools.