Veridect — Fake News & Propaganda Detection Tool

A full-stack fake news and propaganda detection web application built with Flask, vanilla JS, and a custom dark UI. Designed as a final year cybersecurity project.

Features

• NLP Analysis — Detects 9 propaganda tactics via pattern matching (fear appeal, loaded language, false dichotomy, bandwagon, appeal to authority, scapegoating, repetition, black & white thinking, glittering generality)
• Sentiment Scoring — Measures emotional manipulation through linguistic signals
• OSINT Source Analysis — Domain credibility scoring, age detection, registrar country risk, fact-check database lookup
• Multi-signal Fusion — Weighted risk score combining NLP + OSINT signals
• Scan History — Persists all analyses in the session with timeline view
• Propaganda Guide — Reference library of all tactic types with examples
• Domain Lookup — Standalone domain credibility tool
• Responsive Dark UI — Sleek dashboard with animated risk meters

Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Run the server
python app.py

# 3. Open in browser
# http://localhost:5000

Project Structure

veridect/
├── app.py                  # Flask backend + NLP/OSINT analysis engine
├── requirements.txt
├── templates/
│   └── index.html          # Single-page app shell
└── static/
    ├── css/
    │   └── style.css       # Full dark theme stylesheet
    └── js/
        └── app.js          # Frontend logic, API calls, rendering

How It Works

Analysis Pipeline

1. Input — URL or raw article text submitted via the top bar
2. Domain extraction — Parses URL to extract domain for OSINT checks
3. Domain analysis — Scores domain against whitelist, checks naming patterns, TLD, simulates age and registrar country
4. NLP analysis — Regex-based propaganda tactic detection across 9 categories, sentiment scoring via keyword heuristics, manipulation signal extraction
5. Score fusion — Weighted combination: tactics (30%) + sentiment (20%) + manipulation (15%) + source credibility (25%) + cross-source (10%)
6. Output — Risk score, credibility score, per-signal breakdown, OSINT report

Extending to Production

Replace the heuristic components with:

• BERT/RoBERTa fine-tuned on FakeNewsNet or LIAR dataset (HuggingFace Transformers)
• Real Whois API (python-whois library)
• NewsAPI / GDELT for cross-source agreement
• ClaimBuster API for claim-level fact checking
• Google Reverse Image API for image forensics
• PostgreSQL for persistent scan history
• SHAP / LIME for ML explainability

API Endpoints

Method	Endpoint	Description
POST	/api/analyze	Analyse article URL or text. Body: {"text": "..."}
GET	/api/history	Returns sample scan history

Example Request

curl -X POST http://localhost:5000/api/analyze \
  -H "Content-Type: application/json" \
  -d '{"text": "SHOCKING: Government HIDING 5G truth! Share before banned!"}'

Example Response

{
  "input": "SHOCKING: Government HIDING 5G truth!...",
  "scores": {
    "fake_probability": 78,
    "credibility_score": 22,
    "risk_level": "high",
    "source_credibility": 0.5,
    "cross_source_agreement": 0.4
  },
  "domain_analysis": { ... },
  "text_analysis": {
    "tactics": [
      { "key": "fear_appeal", "name": "Fear appeal", "severity": "high" },
      { "key": "loaded_language", "name": "Loaded language", "severity": "high" }
    ],
    "signals": {
      "sentiment_bias": 0.72,
      "linguistic_manipulation": 0.65,
      "claim_verifiability": 0.08
    }
  }
}

Tech Stack

Layer	Technology
Backend	Python, Flask
NLP	Regex heuristics (→ BERT/RoBERTa)
OSINT	Domain pattern analysis (→ Whois, Shodan)
Frontend	Vanilla HTML/CSS/JS, Space Grotesk + DM Mono
Deployment	Python built-in server (→ Docker + Gunicorn)