Image Authenticity Prediction

A comprehensive deep learning framework for predicting image authenticity and understanding model behavior through explainability methods. This project explores how different CNN architectures perceive AI-generated versus real images, with a focus on human-alignment and interpretable AI.

📋 Table of Contents

• Image Authenticity Prediction
  • 📋 Table of Contents
  • 🔍 Overview
  • 📚 Documentation
  • 🚀 Quick Start
    • Installation
    • Dataset Setup
  • 💡 Usage
    • Basic Commands
    • Python API
  • 🤖 Available Models
  • 📊 Dataset
  • 🤝 Contributing
  • 📄 License
  • 🙏 Acknowledgments

🔍 Overview

This research project investigates image authenticity prediction using multiple deep learning architectures, analyzing how different models perceive and evaluate AI-generated content. The work focuses on three key areas:

1. Model Performance & Architecture Comparison

• Evaluates 7 state-of-the-art CNN architectures on authenticity prediction
• Compares traditional supervised (ImageNet) vs. self-supervised (BarlowTwins) pretraining
• Analyzes performance across different network depths and architectural designs

2. Model Explainability & Human Alignment

• Implements GradCAM and Multiscale Pixel Masking for visual explanations
• Compares what different models "look at" when judging authenticity
• Studies alignment between model attention and human perception

3. Network Optimization & Feature Analysis

• Identifies and removes redundant or harmful features through pruning
• Analyzes feature importance across different layers
• Investigates model efficiency and compression possibilities

📚 Documentation

Complete documentation for this project:

• SETUP_GUIDE.md - Setup instructions, dataset structure, import patterns, and troubleshooting
• PROJECT_STRUCTURE.md - Detailed project architecture and module organization
• QUICK_REFERENCE.md - Command reference and common operations
• TODO.md - Development roadmap and known issues
• Experiment Reports:
  • EXPERIMENT_1_TECHNICAL_REPORT.md - Training, pruning, and evaluation
  • EXPERIMENT_2_TECHNICAL_REPORT.md - Explainability methods comparison
  • EXPERIMENT_3_TECHNICAL_REPORT.md - Ensemble learning strategies (WIP)

New to the project? Start with SETUP_GUIDE.md for complete setup instructions.

🚀 Quick Start

Installation

# Install PyTorch with CUDA support
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

# Install dependencies
pip install pandas pillow numpy matplotlib tqdm scipy scikit-learn scikit-image seaborn opencv-python

Dataset Setup

# Create dataset directories
mkdir -p Dataset/AIGCIQA2023 Dataset/Single_score

# Place your data:
# - Dataset/AIGCIQA2023/real_images_annotations.csv
# - Dataset/AIGCIQA2023/Image/ (image files)
# - Dataset/Single_score/ (25 participant CSV files)

See SETUP_GUIDE.md#git-ignored-directories for detailed dataset setup.

💡 Usage

Basic Commands

# Train a model
python -m Image_Authenticity_prediction train --model vgg16 --epochs 50

# Evaluate a model
python -m Image_Authenticity_prediction evaluate --model vgg16 --weights path/to/weights.pth

# Run complete experiments
python -m Image_Authenticity_prediction experiment-one --train --prune --test
python -m Image_Authenticity_prediction experiment-two --xai-methods both
python -m Image_Authenticity_prediction experiment-three --strategy both

Python API

from Image_Authenticity_prediction.main.Models import VGG16AuthenticityPredictor
from Image_Authenticity_prediction.main.data import IMAGENET_DATASET
from Image_Authenticity_prediction.main.train import train_model

# Initialize and train
model = VGG16AuthenticityPredictor(freeze_backbone=True)
# ... training code

For detailed usage, see:

• QUICK_REFERENCE.md - All commands and options
• EXPERIMENT_1_TECHNICAL_REPORT.md - Experiment 1 details
• EXPERIMENT_2_TECHNICAL_REPORT.md - Experiment 2 details

🤖 Available Models

All models use transfer learning with pretrained weights and custom regression heads for authenticity score prediction.

Model	Input Size	Status
VGG16	224×224	✅ Active
VGG19	224×224	✅ Active
ResNet-152	224×224	✅ Active
DenseNet-161	300×300	✅ Active
InceptionV3	299×299	⚠️ Not used in experiments*
EfficientNet-B3	300×300	✅ Active
BarlowTwins	224×224	✅ Active

*InceptionV3 is implemented but excluded from Experiment 1 due to incompatibility with the current pruning method.

📊 Dataset

Uses AIGCIQA2023 dataset with Mean Opinion Scores (MOS) for image authenticity.

Structure:

Dataset/
├── AIGCIQA2023/
│   ├── real_images_annotations.csv    # Aggregated annotations
│   └── Image/                         # Image files
└── Single_score/                      # Individual participant scores (25 CSV)$$

Split: 70% train, 10% validation, 20% test (seed: 42)

🤝 Contributing

Research project by Icaro Re Depaolini as part of thesis work at CiMEC, University of Trento.

For questions or contributions, please contact the project maintainer.

📄 License

Academic research project. Contact authors for usage permissions.

🙏 Acknowledgments

• Pretrained models from PyTorch and torchvision
• BarlowTwins implementation from Facebook Research
• AIGCIQA2023 dataset authors
• CiMEC, University of Trento

---

Author: Icaro Re Depaolini
Institution: CiMEC, University of Trento
Last Updated: January 2026