VeriStream - Video Authenticity Verification Platform

State-of-the-art forensic analysis for video content verification

Features • Quick Start • Architecture • Documentation

---

Overview

VeriStream is an advanced video forensic analysis platform that uses 8 distinct AI-powered vectors to assess video authenticity. It combines computer vision, signal processing, biometric analysis, and semantic reasoning to detect deepfakes, splicing, and other manipulations.

Why VeriStream?

In an era of AI-generated content and sophisticated video manipulation, verifying media authenticity is critical:

• Misinformation Prevention: Combat fake news and deepfakes in journalism and social media
• Legal Evidence: Verify video evidence integrity for court proceedings
• Content Authentication: Validate user-generated content for platforms
• Corporate Security: Detect manipulated videos in corporate communications
• Digital Forensics: Investigate potential video tampering in forensic cases

Key Features

✅ Deepfake Detection - ViT-based AI model
✅ Editorial Integrity - Cut detection & AV sync analysis
✅ Semantic Accuracy - Whisper transcription + CLIP + LLM reasoning
✅ Provenance Verification - Metadata analysis & C2PA validation
✅ Biological Consistency - Blink rate & rPPG heart rate analysis
✅ Physical Laws - RAFT optical flow & lighting analysis
✅ Acoustic Physics - YAMNet environmental sound classification
✅ Adversarial Defense - JPEG artifacts & frequency analysis

---

User Interface

Landing Page

Video Upload

Analysis Results

Detailed Report

---

The 8 Forensic Vectors

Vector A: Deepfake Detection (30%)

Technology: Vision Transformer (ViT) deep learning model
What it detects: AI-generated faces, synthetic artifacts, GAN fingerprints
Use Cases:

• Detecting face-swap deepfakes in political videos
• Identifying AI-generated synthetic media
• Validating authentic source footage

Vector B: Editorial Integrity (20%)

Technology: PySceneDetect, MediaPipe Face Mesh, Audio-Video Correlation
What it detects: Scene cuts, splicing, audio-video desynchronization
Use Cases:

• Identifying edited interview clips taken out of context
• Detecting spliced footage in news reports
• Verifying continuous recording in surveillance videos

Vector C: Semantic Accuracy (15%)

Technology: Whisper (speech-to-text), CLIP (vision-language), Ollama LLM
What it detects: Audio-visual mismatches, narrative inconsistencies
Use Cases:

• Detecting dubbed or voice-swapped content
• Identifying mismatched audio overlays
• Validating that visuals match spoken content

Vector D: Provenance & Metadata (10%)

Technology: FFprobe, ExifTool, C2PA Cryptographic Verification
What it detects: Missing metadata, timestamp tampering, chain of custody
Use Cases:

• Verifying camera origin and capture device
• Detecting re-encoded or re-compressed videos
• Validating Content Credentials (C2PA) for professional content

Vector E: Biological Consistency (10%)

Technology: MediaPipe landmarks, Eye Aspect Ratio (EAR), CHROM rPPG
What it detects: Unnatural blink rates, synthetic pulse patterns
Use Cases:

• Detecting face-replacement deepfakes
• Identifying AI-generated human faces
• Validating physiological realism

Vector F: Physical Laws (8%)

Technology: RAFT Optical Flow, Depth Estimation, Lighting Analysis
What it detects: Impossible physics, lighting inconsistencies, motion artifacts
Use Cases:

• Detecting object insertion/removal
• Identifying unrealistic motion patterns
• Verifying consistent environmental lighting

Vector G: Acoustic Physics (5%)

Technology: YAMNet (environmental sound classification), Spectral Analysis
What it detects: Background sound inconsistencies, audio splicing
Use Cases:

• Detecting acoustic environment changes (indoor/outdoor jumps)
• Identifying replaced audio segments
• Validating continuous acoustic signatures

Vector H: Adversarial Defense (2%)

Technology: JPEG artifact analysis, FFT frequency domain, Re-compression tests
What it detects: Anti-forensic tampering attempts
Use Cases:

• Detecting adversarial noise injection
• Identifying compression-based obfuscation
• Uncovering sophisticated tampering attempts

---

Quick Start with Docker

Prerequisites

• Docker (v20.10+) - Install Docker
• Docker Compose (v2.0+) - Install Docker Compose
• 8GB+ RAM recommended for model loading
• (Optional) NVIDIA GPU with Docker GPU support for Ollama LLM

1. Clone the Repository

git clone https://github.com/yourusername/VeriStream.git
cd VeriStream

2. Build and Run with Docker Compose

# Build and start all services
docker-compose up --build -d

# Check service status
docker-compose ps

# View logs
docker-compose logs -f

3. Access the Application

• Frontend: http://localhost
• Backend API: http://localhost:8000
• API Docs: http://localhost:8000/docs

4. Upload and Analyze

1. Open http://localhost in your browser
2. Click "Upload Video" or drag & drop
3. Wait for analysis (typically 40-80 seconds)
4. View detailed forensic report and download verified video

---

Docker Services

The application consists of three services:

Backend (Python FastAPI)

• Port: 8000
• Tech: FastAPI, PyTorch, MediaPipe, Whisper, CLIP
• Volumes:
  • backend_temp - Temporary processing files
  • backend_weights - Downloaded model weights
  • backend_database - Analysis history database

Frontend (React + Nginx)

• Port: 80
• Tech: React 18, Vite, TailwindCSS
• Features: Real-time progress, responsive UI, PDF export

Ollama (Optional LLM Service)

• Port: 11434
• Requirements: NVIDIA GPU recommended
• Enable: Uncomment the ollama service in docker-compose.yml

---

Environment Variables

Create a .env file in the root directory (optional):

# Ollama LLM endpoint (if running separately)
OLLAMA_BASE_URL=http://ollama:11434

# Backend development mode
RELOAD=true

---

Development Mode

For local development with hot-reload:

Backend

cd veristream_backend
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn main:app --reload --host 0.0.0.0 --port 8000

Frontend

cd veristream_frontend
npm install
npm run dev

---

Architecture

┌──────────────────────────────────────────────────────────────────────┐
│                         VeriStream Platform                          │
├──────────────────────────────────────────────────────────────────────┤
│                                                                      │
│   ┌──────────────┐         ┌───────────────┐        ┌─────────────┐  │
│   │   Frontend   │────────▶│    Backend    │◀──────▶│   Models    │  │
│   │ React + Nginx│         │    FastAPI    │        │   Weights   │  │
│   │   Port 80    │         │   Port 8000   │        │    Cache    │  │
│   └──────────────┘         └───────┬───────┘        └─────────────┘  │
│                                    │                                 │
│   ┌────────────────────────────────┴────────────────────────────-┐   │
│   │              8 Forensic Analysis Vectors                     │   │
│   ├──────────────────────────────────────────────────────────────┤   │
│   │  A. Deepfake Detection        (ViT Model)                    │   │
│   │  B. Editorial Integrity       (PySceneDetect, AV Sync)       │   │
│   │  C. Semantic Accuracy         (Whisper, CLIP, Ollama)        │   │
│   │  D. Provenance & Metadata     (C2PA Verification)            │   │
│   │  E. Biological Consistency    (rPPG, Blink Analysis)         │   │
│   │  F. Physical Laws             (RAFT Optical Flow)            │   │
│   │  G. Acoustic Physics          (YAMNet)                       │   │
│   │  H. Adversarial Defense       (FFT, JPEG Artifacts)          │   │
│   └──────────────────────────────────────────────────────────────┘   │
│                                                                      │
│   ┌──────────────────────────────────────────────────────────────┐   │
│   │                    Persistent Volumes                        │   │
│   ├──────────────────────────────────────────────────────────────┤   │
│   │  • backend_temp        (Temporary processing files)          │   │
│   │  • backend_weights     (Downloaded model cache)              │   │
│   │  • backend_database    (Analysis history database)           │   │
│   └──────────────────────────────────────────────────────────────┘   │
└──────────────────────────────────────────────────────────────────────┘

---

Documentation

• Backend Flow - Step-by-step execution flow
• Forensic Logic - Detailed scoring algorithm
• API Documentation - Interactive Swagger UI

---

Troubleshooting

Container won't start

# Check logs
docker-compose logs backend
docker-compose logs frontend

# Rebuild from scratch
docker-compose down -v
docker-compose up --build

"Out of memory" errors

Increase Docker memory allocation:

• Docker Desktop: Settings → Resources → Memory (set to 8GB+)

Models not downloading

Models are downloaded on first use. Ensure:

• Internet connection is active
• HuggingFace is accessible
• Sufficient disk space (5GB+ for models)

Ollama LLM not working

# Install and pull model
docker exec -it veristream-ollama ollama pull llama3.1:8b

# Verify
curl http://localhost:11434/api/tags

Frontend shows "Cannot connect to backend"

1. Verify backend is healthy: docker-compose ps
2. Check backend logs: docker-compose logs backend
3. Ensure nginx proxy is configured correctly

---

Performance

Module	Typical Duration	Parallelizable
Deepfake Detection	2-5s	✅
Editorial Integrity	5-10s	Partial
Semantic Accuracy	10-20s	✅
Provenance	1-2s	✅
Biological	8-15s	✅
Physical Laws	5-10s	✅
Acoustic Physics	8-12s	✅
Adversarial Defense	3-5s	✅

Total: ~40-80 seconds for a 30-second video (sequential)

---

Made with ❤️ for video authenticity verification

⭐ Star us on GitHub if this project helped you!