A Novel Deep Federated Learning-Based Model to Enhance Privacy in Critical Infrastructure Systems

Abstract

Deep learning (DL) can provide critical infrastructure operators with valuable insights and predictive capabilities to help them make more informed decisions, improving system's robustness. However, training DL models requires large amounts of data, which can be costly to store in a centralized manner. Storing large amounts of sensitive critical infrastructure data in the cloud can pose significant security risks. Federated learning (FL) allows several clients to share learning data and train ML models. Unlike centralized models, FL does not require the sharing of client data. A novel framework is presented to train a VGG16 based CNN global model without sharing the data and only updating the local models among clients using federated averaging. For experimentation, MNIST dataset is used. The framework achieves high accuracy and keep data private using FL in critical infrastructures. The benefits and challenges of FL along with security vulnerabilities and attacks have been discussed along with the defenses that can be used to mitigate these attacks.