Differential Privacy-Preserving Algorithms for Secure Training of Machine Learning Models

Since ML is used more widely in data-driven apps, issues about data privacy and protection are becoming more common. It provides a method for assessing and controlling the privacy of individuals in datasets used for machine learning (ML) training. This paper examines DP-preserving algorithms designed for the safe training of machine learning models. We study centralized, local, and distributed methods for applying differential privacy to the training of logistic regression, support vector machines, and deep neural networks. Next, we investigate the foundations of differential privacy, including privacy budgets, the concept of sensitive data, and noise addition, and examine how they impact the accuracy and reliability of the model. We apply DP-SGD, examine its effects on utility and privacy, and study models that combine federated learning with secure multi-party computation. We utilize the MNIST, CIFAR-10, and Adult Income datasets in a comprehensive experiment to evaluate the accuracy, privacy loss, convergence, and runtime of our system. While training a DP model incurs costs in utility, our testing shows that selecting the right parameters and utilizing a combination of privacy approaches can yield secure and high-performing results. Our research aims to inform machine learning (ML) research on privacy issues and provide guidance on implementing differential privacy in ML applications