📰 VeriFy — Fake News Detection & Explanation System

VeriFy is an intelligent Fake News Detection and Explanation system that combines deep learning–based text classification with LLM-powered explainability.
The project aims to not only classify news as Fake or Real, but also explain why a particular decision was made and collect user feedback to improve future model training.

---

🚀 Project Motivation

With the rapid spread of misinformation across digital platforms, users often struggle to judge the credibility of online news.
Most existing tools either:

• Only classify news without explanation, or
• Provide generic AI-generated responses without grounding in model behavior.

VeriFy bridges this gap by combining:

• A trained deep learning model for classification
• Explainable AI (XAI) using Gemini
• Counterfactual reasoning
• Human-in-the-loop feedback collection

---

🧠 How the System Works

1️⃣ Text Classification (Core Model)

• News text is preprocessed and converted into sequences using a Tokenizer
• The sequences are padded to a fixed length (maxlen = 1000)
• A Word2Vec + LSTM neural network predicts whether the news is Fake or Real

2️⃣ Explainability (LLM Integration)

• The prediction and original text are passed to Gemini LLM
• Gemini generates human-readable reasoning explaining why the model may have made that prediction
• This improves transparency and user trust

3️⃣ Counterfactual Analysis

• The system highlights how small wording changes could flip the prediction
• Helps identify sensational terms, emotional language, or misleading phrases

4️⃣ User Feedback Loop

• Users can confirm whether the prediction was correct
• Feedback is logged into a CSV file
• This data can be reused for future re-training and dataset expansion

---

🏗️ Model Architecture

• Embedding Layer

  • Pre-trained Word2Vec embeddings (100 dimensions)
  • Embeddings frozen for stable semantic learning

• LSTM Layer

  • Captures long-term dependencies in news articles
  • Effective for modeling narrative flow and context

• Dense Output Layer

  • Sigmoid activation for binary classification (Fake / Real)

📌 Why LSTM + Word2Vec?

• Handles long news articles better than traditional ML models
• Learns semantic and contextual patterns
• Performs well on sequential text data
• More interpretable than large black-box transformer models for this task

---

🧪 Dataset

• Source: Aggregated fake and real news articles
• Preprocessing:
  • Special character removal
  • Tokenization
  • Padding to uniform sequence length
• Labels:
  • 0 → Fake
  • 1 → Real

---

🖥️ Tech Stack

Core Technologies

• Python
• TensorFlow / Keras
• Gensim (Word2Vec)
• Streamlit (Frontend UI)
• Google Gemini API (Explainability)
• Pandas / NumPy

Explainability & Feedback

• Gemini LLM for explanations
• CSV-based feedback logging for retraining

---

---

🔗 Model & Tokenizer Download

Due to GitHub size limits, the trained model and tokenizer are hosted on Google Drive.

📥 Download from here:
👉 Google Drive Link – Model & Tokenizer

After downloading:

• Place my_model.h5 inside model/
• Place tokenizer.pkl inside model/

---

⚙️ Setup Instructions

1️⃣ Clone the Repository

git clone https://github.com/your-username/verify-fake-news.git
cd verify-fake-news

2️⃣ Create Virtual Environment (Recommended)

python -m venv venv

Activate on Linux / macOS

source venv/bin/activate

Activate on Windows

venv\Scripts\activate

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Set Gemini API Key

Create a file named api_key.env in the project root and add:

GEMINI_API_KEY=your_api_key_here

▶️ Run the Application