AI vs Real Image and Video Classification

MLOps project for binary classification of AI-generated vs real images and videos using deep learning.

Introduction

The proliferation of AI-generated content has created a need for automated detection systems. This project implements a deep learning solution to distinguish between authentic and AI-generated images and videos, addressing challenges in content verification and media authenticity.

Team

• Muhammad Basil
• Taha Zahid

MLOps Stack

• Airflow - Workflow orchestration
• DagsHub - Data versioning and collaboration
• MLflow - Experiment tracking and model registry
• MongoDB - Data storage
• Git - Version control
• Deep Learning - Model training

Technology

• EfficientNet - Transfer learning model architecture
• FFmpeg - Video frame extraction
• TensorFlow/Keras - Deep learning framework

Proposed Methodology

1. Data Preprocessing

   • Extract video frames using FFmpeg
   • Clean and resize images to 256x256
   • Balance dataset classes
   • Split into train/validation/test sets

2. Model Architecture

   • EfficientNetB0 with transfer learning (ImageNet weights)
   • Two-phase training: frozen backbone then fine-tuning
   • Data augmentation: rotation, flipping, zoom, brightness, contrast

3. MLOps Pipeline

   • Airflow orchestrates preprocessing and training workflows
   • MLflow tracks experiments and model versions
   • DagsHub manages data versioning
   • MongoDB stores metadata and results
   • Git for code version control

4. Training Process

   • Phase 1: Train with frozen EfficientNetB0 backbone (10 epochs)
   • Phase 2: Fine-tune last 20 layers (10 epochs)
   • Callbacks: ModelCheckpoint, EarlyStopping, ReduceLROnPlateau

Initial Results

• Test Accuracy: 86%
• Model successfully distinguishes between real and AI-generated content
• EfficientNetB0 transfer learning approach shows strong performance

Code

Training code available in:

• train.py - Main training script
• train.ipynb - Jupyter notebook with full pipeline

Key components:

• Data preprocessing and augmentation
• EfficientNetB0 model implementation
• Training and evaluation loops
• Model saving and inference

Project Structure

mlops/
├── train.py
├── train.ipynb
└── README.md

Dataset

Expected structure:

data/
├── train/real/
├── train/fake/
├── val/real/
├── val/fake/
├── test/real/
└── test/fake/

Videos are processed using FFmpeg to extract frames before training.

Usage

python train.py

Or use the Jupyter notebook train.ipynb.

Requirements

pip install tensorflow keras numpy pillow matplotlib

Note: Originally developed in Google Colab. Update paths and remove Colab-specific imports for local execution.