README.md

metadata

title: Image-Forensic
emoji: 🔍
colorFrom: indigo
colorTo: purple
sdk: docker
pinned: false

ImageForensics-Detect

Research-grade multi-branch image forensics platform for detecting Real vs. AI-Generated images.
B.Tech Final Year Project · IEEE-Style Research System

---

🧠 What This Does

ImageForensics-Detect analyzes uploaded images through 5 independent forensic detection branches and fuses their outputs using certainty-weighted probabilistic fusion to decide whether an image is:

• Real — captured by a physical camera
• AI-Generated — created by GANs or diffusion models (Stable Diffusion, DALL-E, Midjourney, etc.)

---

🏗️ Architecture

Input Image
    │
    ├──► Spectral Branch   (FFT/DCT analysis)         [no training needed]
    ├──► Edge Branch       (Sobel/Laplacian forensics) [no training needed]
    ├──► CNN Branch        (EfficientNet-B0 / TF)      [train_cnn.py]
    ├──► ViT Branch        (ViT-B/16 / PyTorch+timm)   [train_vit.py]
    └──► Diffusion Branch  (residual noise analysis)   [no training needed]
             │
    Certainty-Weighted Probabilistic Fusion
             │
    ┌────────────────────────────────┐
    │  Prediction: Real / AI-Gen     │
    │  Confidence: 97.1%             │
    │  Grad-CAM heatmap              │
    │  Spectral anomaly map          │
    │  Noise residual map            │
    └────────────────────────────────┘

---

📁 Folder Structure

ImageForensics-Detect/
├── data/raw/{real,fake}/          ← Your dataset goes here
├── models/                        ← Saved .h5 / .pth weights
├── branches/
│   ├── spectral_branch.py         ✅ COMPLETE (signal processing)
│   ├── edge_branch.py             ✅ COMPLETE (signal processing)
│   ├── cnn_branch.py              🔵 BASELINE (needs training)
│   ├── vit_branch.py              🔵 BASELINE (needs training)
│   └── diffusion_branch.py        ✅ COMPLETE (signal processing)
├── fusion/fusion.py               ✅ COMPLETE
├── explainability/
│   ├── gradcam.py                 ✅ COMPLETE
│   └── spectral_heatmap.py        ✅ COMPLETE
├── training/
│   ├── dataset_loader.py          ✅ COMPLETE
│   ├── train_cnn.py               ✅ COMPLETE
│   ├── train_vit.py               ✅ COMPLETE
│   └── evaluate.py                ✅ COMPLETE
├── backend/app.py                 ✅ COMPLETE (FastAPI)
├── frontend/{index.html,style.css,app.js}  ✅ COMPLETE
├── utils/{image_utils.py,logger.py}        ✅ COMPLETE
└── outputs/                       ← Logs, heatmaps, eval results

---

⚙️ Installation

Prerequisites

• Python 3.9+
• pip / conda

1. Create Virtual Environment

cd "ImageForensics-Detect"
python -m venv venv
source venv/bin/activate          # macOS/Linux
# venv\Scripts\activate           # Windows

2. Install Dependencies

pip install -r requirements.txt

Note: On Apple Silicon (M1/M2/M3), use pip install tensorflow-macos tensorflow-metal instead of tensorflow.

---

🗂️ Dataset Setup

Populate the dataset folders before training:

data/raw/
├── real/    ← Real camera photos (.jpg, .png)
└── fake/    ← AI-generated images (.jpg, .png)

Recommended datasets:

Type	Dataset	Source
Real	RAISE-1K / VISION / MIT-5k	Kaggle / research groups
AI-Gen	ThisPersonDoesNotExist / SDXL outputs	Collected/scraped
Mixed	ArtiFact / CNNDetection	GitHub papers

The loader auto-splits: 70% train / 15% val / 15% test (stratified).

---

🏋️ Training

Train CNN Branch (EfficientNet-B0 / TensorFlow)

python training/train_cnn.py --epochs 30 --batch_size 32 --lr 1e-4
# Saves: models/cnn_branch.h5

Train ViT Branch (ViT-B/16 / PyTorch)

python training/train_vit.py --epochs 20 --batch_size 16 --lr 1e-4
# Saves: models/vit_branch.pth

Without training: The system is still functional — the 3 handcrafted branches (Spectral, Edge, Diffusion) produce real forensic outputs immediately. CNN/ViT branches return neutral 0.5 confidence and are flagged as "untrained" in the API response.

---

📊 Evaluation

# Evaluate entire fusion system
python training/evaluate.py

# Evaluate individual branches
python training/evaluate.py --branch spectral
python training/evaluate.py --branch edge
python training/evaluate.py --branch cnn
python training/evaluate.py --branch vit
python training/evaluate.py --branch diffusion

Reports saved to outputs/:

• confusion_matrix_<branch>.png
• roc_curve_<branch>.png
• evaluation_<branch>.csv

---

🚀 Running the System

Step 1: Start Backend API

uvicorn backend.app:app --reload --host 0.0.0.0 --port 8000

Step 2: Open Frontend

Open frontend/index.html in your browser (double-click, or use Live Server in VS Code).

Step 3: Upload and Analyze

Drag-and-drop any image → Click Analyze Image → View results.

---

🌐 API Reference

POST /predict

Upload an image and receive full forensic analysis.

Request:

curl -X POST "http://localhost:8000/predict" \
     -F "file=@your_image.jpg"

Response:

{
  "prediction": "AI-Generated",
  "confidence": 97.1,
  "prob_fake": 0.9855,
  "branches": {
    "spectral":  { "prob_fake": 0.9420, "confidence": 0.8800, "label": "AI-Generated" },
    "edge":      { "prob_fake": 0.8100, "confidence": 0.7200, "label": "AI-Generated" },
    "cnn":       { "prob_fake": 0.9820, "confidence": 0.9640, "label": "AI-Generated" },
    "vit":       { "prob_fake": 0.9600, "confidence": 0.9200, "label": "AI-Generated" },
    "diffusion": { "prob_fake": 0.8900, "confidence": 0.8300, "label": "AI-Generated" }
  },
  "gradcam_b64":   "<base64-encoded JPEG>",
  "spectrum_b64":  "<base64-encoded JPEG>",
  "noise_map_b64": "<base64-encoded JPEG>",
  "edge_map_b64":  "<base64-encoded JPEG>",
  "low_certainty": false
}

GET /health

{ "status": "ok", "service": "ImageForensics-Detect", "version": "1.0.0" }

GET /logs

{ "total": 42, "real": 18, "ai_generated": 24 }

---

📝 Research Methodology

Title (Suggested)

"Multi-Branch Certainty-Weighted Forensic Detection of AI-Generated Images Using Spectral Analysis, Edge Statistics, CNN, and Vision Transformers"

Abstract

This work presents ImageForensics-Detect, a multi-branch forensic analysis framework for distinguishing real camera photographs from AI-generated images produced by GANs and diffusion models. The system integrates five complementary detection branches: (1) spectral analysis using FFT and DCT to capture frequency-domain artifacts; (2) edge analysis using Sobel/Laplacian operators and gradient distribution statistics; (3) a CNN branch (EfficientNet-B0) for local texture and patch-level artifact detection; (4) a ViT branch (ViT-B/16) for global semantic inconsistency detection; and (5) a diffusion residual branch analyzing noise kurtosis and spatial uniformity. Branch predictions are combined using certainty-weighted probabilistic fusion, ensuring that uncertain or untrained branches contribute proportionally less to the final decision.

Key Design Decisions

Decision	Rationale
Multi-branch ensemble	No single signal catches all generator types
Certainty-weighted fusion	Prevents weak/untrained branches from degrading accuracy
FFT + DCT (spectral)	GAN checkerboard artifacts are frequency-domain detectable
EfficientNet-B0	Best accuracy-efficiency trade-off for the CNN branch
ViT-B/16 (timm)	Global receptive field catches semantic incoherence CNNs miss
Noise uniformity (diffusion)	Diffusion models produce spatially uniform noise fields

References (Key Papers)

1. Wang et al. (2020). CNN-generated images are surprisingly easy to spot... for now. CVPR.
2. Frank et al. (2020). Leveraging Frequency Analysis for Deep Fake Image Recognition. ICML.
3. Corvi et al. (2023). On the detection of synthetic images generated by diffusion models. ICASSP.
4. Ojha et al. (2023). Towards Universal Fake Image Detection. CVPR.
5. Selvaraju et al. (2017). Grad-CAM: Visual Explanations from Deep Networks. ICCV.

---

🛠️ Tech Stack

Layer	Technology
CNN Branch	TensorFlow 2.13+, EfficientNet-B0
ViT Branch	PyTorch 2.0+, timm, ViT-B/16
Signal Branches	OpenCV, NumPy, SciPy
Backend	FastAPI, Uvicorn
Frontend	HTML5, CSS3, Vanilla JavaScript
Evaluation	Scikit-learn, Matplotlib, Seaborn

---

📌 Status Summary

Module	Status	Result
Spectral Branch	✅ Complete	Signal Forensics
Edge Branch	✅ Complete	Signal Forensics
Diffusion Branch	✅ Complete	Signal Forensics
CNN Branch	✅ Trained	EfficientNet-B0
ViT Branch	✅ Fully Trained	99.30% Accuracy
Fusion Module	✅ Complete	Certainty-Weighted
Grad-CAM	✅ Complete	(with Saliency Fallback)
Frontend UI	✅ Enhanced	Stats Hero + Prob Bar
FastAPI Backend	✅ Complete	Port 8000

---

Built for IEEE-style research deployment. B.Tech Final Year Project.