Quantifying uncertainty in adversarial federated learning

Abstract

The research project called Quantifying Uncertainty in Adversarial Federated Learning aims to analyze and propose new approaches to distributed machine learning models that maintain privacy and security restrictions. Federated Learning (FL) is a promising approach to training data collaboratively on distributed devices while accounting for privacy restrictions. However, the FL training process is vulnerable to model poisoning attacks where malicious participants can upload fake model weights. The project aims to address these vulnerabilities and propose new solutions for maintaining privacy and security in distributed machine learning models. In short, this project presents a scientific research proposal in five directions: (i) quantification of model generalization based on Bayesian neural networks for federated learning systems; (ii) DDoS intrusion detection system approaches in federated applications; (iii) uncertainty quantification in distributed heterogeneous environment (e.g., Federated Learning); (iv) investigation for continual (incremental) learning to identify unknown new malware is necessary to protect systems even at day zero of a malware release; and (v) study the use of ordinal patterns statistical tests to identify data poisoning attacks in federated applications. (AU)