Degree

Master of Science in Data Science

Department

Department of Computer Science

Faculty/ School

School of Mathematics and Computer Science (SMCS)

Date of Submission

Fall 2024

Supervisor

Ms. Tasbiha Fatima, Lecturer, Department of Computer Science

Keywords

Federated Learning, Differential Privacy, MNIST, Deep Learning, Gradient Clipping

Abstract

This project explores the application of federated learning combined with differential privacy (DP) for training a deep learning model on the MNIST dataset. Federated learning enables distributed training across multiple clients while keeping data decentralized, which is essential for privacy preservation. The introduction of differential privacy further ensures that individual client data remains protected by adding noise to the model's gradients during training, thereby limiting the exposure of sensitive information.

The code implemented in this project focuses on setting up a federated learning framework and employing a deep neural network for classification of digits from the MNIST dataset. Parameters of differential privacy are added including noise multipliers and gradient clipping values to ensure preservation of user data.

The dataset is divided into three partitions, each excluding certain digit groups to simulate non-IID (non-independent and identically distributed) data scenarios. Federated averaging strategy is applied for global model updates.

Evaluation metrics, including model accuracy and confusion matrices, are computed after each round of training. The results indicate that the proposed approach can maintain high model performance while respecting differential privacy constraints.

Overall, this project demonstrates how federated learning with differential privacy can provide a practical solution for training machine learning models securely across distributed datasets, paving the way for privacy-preserving applications in machine learning.

Document Type

Restricted Access

Submission Type

Research Project

The full text of this document is only accessible to authorized users.

Share

COinS