FatimahEmadEldin/GenAI-RealEstate-Detector

🕵️‍♀️ GenAI Real Estate Forensic Detector (Track B)

Status: 🏆 Competition Submission for MenaML Winter School 2026
Track: B (Real Estate & Commercial Integrity)
Model Architecture: EfficientNetB0 (Fine-tuned)
Framework: TensorFlow / Keras

📝 Project Overview

As Generative AI becomes mainstream, "virtual staging" in real estate can mislead buyers by hiding structural flaws or fabricating non-existent features.

This project provides a Forensic Signal Detector capable of identifying GenAI-generated or heavily manipulated images in high-stakes public domains. It focuses on detecting:

1. Texture Consistency: "Unnatural smoothness" in walls/skin.
2. Compression Discrepancies: Digital splicing of objects.
3. Frequency Domain Artifacts: Mathematical fingerprints left by GANs/Diffusion models.

📊 Evaluation Results

The model was trained on a dynamic stream of Places365 images with synthetic artifact injection.

Metric	Best Value	Epoch Achieved
Training Accuracy	99.07%	Epoch 1
Training Loss	0.0201	Epoch 1
Validation Stability	97.34%	Epoch 5

📦 Repository Contents

File	Description
forensic_model.h5	The Model. Fine-tuned EfficientNetB0 weights.
predict.py	Inference Script. Automates the detection process for a folder of images.
hackathon_test_set.zip	The Dataset. Contains 1,000 images (500 Real / 500 Fake) for testing.
requirements.txt	Dependencies. Required Python libraries.

💻 How to Run Inference

1. Installation

Clone the repository and install dependencies:

git clone [https://huggingface.co/](https://huggingface.co/)[FatimahEmadEldin]/{GenAI-RealEstate-Detector}
cd {GenAI-RealEstate-Detector}
pip install -r requirements.txt

2. Run Prediction

You can run the detection script on any folder of images:

python predict.py --input_dir ./path_to_images --output_file predictions.json

3. Output Format

The system produces a JSON report with authenticity scores and VLM-based reasoning:

{{
  "image_name": "living_room_01.jpg",
  "authenticity_score": 0.99,
  "manipulation_type": "Synthetic/Inpainting",
  "vlm_reasoning": "Forensic analysis detected high-frequency artifacts (0.99). Visual context suggests issues near: window reflection."
}}

🧠 Training Strategy: "Self-Supervised Artifact Injection"

Instead of relying on a static dataset of "Fake" images (which may become outdated), we implemented an On-the-Fly Artifact Generator that corrupts real images (from Places365) during training.

The 3 Forensic Signals:

1. Simulating Upscalers: Applied randomized Bi-cubic downsampling + Gaussian Blur.
2. Simulating Diffusion: Injected High-Frequency Gaussian Noise distributions.
3. Simulating GANs: Injected subtle Checkerboard Grid patterns (frequency domain fingerprints).

📄 License

This project uses the Places365 dataset (MIT CSAIL). The code is open-source under the MIT License.