Projects
- Adversarially Robust Federated Learning Leveraging Confidence Score to Identify Retinal Disease
My undergraduate thesis, supervised by Dr. Zavid Parvez and Tanzim Reza, addresses the challenge of Byzantine attacks in federated learning, where malicious clients can undermine the global model. We developed a method that uses confidence scores to minimize the impact of these harmful updates while maintaining model accuracy. We tested the approach on a retinal OCT dataset, including images of age-related macular degeneration and diabetic macular edema, and saw significant improvements in precision, recall, and F1 score for both InceptionV3 and VGG19. Our method offers a promising solution for improving the security and performance of federated learning models, especially in privacy-sensitive areas like healthcare
- Encrypted Federated Learning for Detecting Skin Lesions
Working under the guidance of Tanzim Reza, I explored various image augmentation methods, including SMOTE, to address class imbalance issues in skin lesion datasets. Additionally, I implemented a BFV-based homomorphic encryption system within the federated learning framework, ensuring privacy-preserving training across distributed data sources while maintaining data security and model performance.
- Malaria-Infected Cell Classifier using Vision Transformer
Implemented the Vision Transformer architecture, achieving 95.48% accuracy in classifying malaria-infected cells using the Vision Transformer model.
- Explainable Neural Network for Alzheimer’s Disease Analysis
Trained a classifier using InceptionV3 and VGG19 architectures and incorporated regional explanations using the LIME technique to improve model interpretability for Alzheimer’s disease analysis.