dr-screening-tool

Diabetic Retinopathy Screening - 3-Model Ensemble + FL Defense

A lightweight, privacy-preserving DR severity classification system with an integrated Federated Learning defense mechanism. MS Thesis project.

Models Included

Each model is a frozen pretrained backbone + custom classifier head: GAP Dense(256,ReLU) BN Dropout(0.5) Dense(128,ReLU) BN Dropout(0.3) Dense(64,ReLU) Dense(5,Softmax)

Classification Classes

Ensemble + Calibration

The three models are combined via probability averaging with temperature scaling (T=1.5, calibrated on the APTOS 2019 validation set to minimize negative log-likelihood).

Test accuracy: ~74% on a combined APTOS 2019 + IDRiD subsample (2,000 images).

How to Run Locally

pip install -r requirements.txt
python app.py

The Gradio app launches at http://localhost:7860 with:

• Multi-image upload with automatic fundus validation
• DR severity classification with confidence scores
• Grad-CAM explainability heatmaps
• Downloadable clinical HTML report
• FL Defense visualization section

Dataset

• APTOS 2019 Blindness Detection (3,662 images) - Kaggle
• IDRiD (516 images) - IEEE ISBI 2019

Federated Learning Defense

This repo also contains code for evaluating a lightweight FL defense against label-flipping poisoning attacks. The defense uses:

1. Output-layer gradient delta extraction (325-dimensional vectors)
2. Pairwise cosine similarity + K-Means clustering (k=2)
3. Temporal consistency tracking with permanent blacklisting
4. Combined Defense ensemble (K-Means + Public Validation + FoolsGold-style consensus)

See the training script train_ensemble_colab.py for the full pipeline.

Citation

@mastersthesis{haris2026dr,
  title={A Lightweight Defense Against Label-Flipping Poisoning Attacks in Federated Learning for Diabetic Retinopathy Detection},
  author={Haris},
  year={2026},
  school={COMSATS University Islamabad, Abbottabad Campus}
}