AI-Generated Image Recognition Project on pyTorch

This project implements Machine Learning pipeline to detect whether an image is real or AI-generated. It features Backend API serving multiple trained models (two CNN-based models and ResNet-50 model) and Frontend dashboard for interactive image analysis and prediction

Table of Contents

• Project Overview
• Project Demonstration
• Dataset
• Usage
• Model Results
  • Benchmarks
  • Baseline CNN Results
  • Tuned CNN Results
  • ResNet-50 Results

---

Project Overview

The goal of this application is to classify images as "REAL" or "FAKE"

Key Features:

• Multiple Models: Comparison between Baseline CNN, Hyperparameter Tuned CNN, ResNet-50 models
• Backend: FastAPI service that loads pre-trained models and handles inference requests
• Frontend: Streamlit application allowing users to upload images and view classification confidence across models
• Analysis: Visualizations of training metrics and model efficiency

---

Project Demonstration

Here is how the application looks and functions:

1. Interactive Detection Dashboard

User can upload image to analyze it using the ensemble of models. The system displays the prediction (Real/Fake) and confidence level for each model

2. Training Metrics

A dedicated tab displays the comparative performance (Precision, Recall, F1-Score) and training time efficiency of the models.

---

Dataset

The models were trained using the CIFAKE Dataset (Real vs. AI-Generated Synthetic Images): Link to CIFAKE Dataset

---

Usage

• Python 3.10+

1. Download libraries

pip install -r requirements.txt 

2. Run the Backend

uvicorn backend.main:app -reload

The API will be available at http://127.0.0.1:8000.

3. Run the Frontend

streamlit run frontend/app.py

The interface will open in your browser at http://localhost:8501.

---

Model Results

We evaluated three different architectures. Below are the visual benchmarks and detailed statistics.

Benchmarks

Important Observation: Although the Baseline CNN and Tuned CNN have close evaluation metrics, in practice ResNet-50 recognizes AI-generated images better and proves to be more robust in real-world scenarios.

Baseline CNN Results

============================================================
                        TEST RESULTS                        
============================================================
- Precision:    0.9256
- Recall:       0.9614
- F1 Score:     0.9432
- Training Time: ~375s

Tuned CNN Results

============================================================
                        TEST RESULTS                        
============================================================
- Precision:    0.9513
- Recall:       0.9413
- F1 Score:     0.9463
- Training Time: ~1290s

ResNet-50 Results

============================================================
                        TEST RESULTS                        
============================================================
- Precision:    0.9621
- Recall:       0.9591
- F1 Score:     0.9606
- Training Time: ~4878s