Faculty Mentor Information
Dr. Hao Chen (Mentor), Boise State University
Abstract
Technology has had a significant influence on the lives of today’s generation. However, over the years, protecting private information while maintaining its utility has become a major concern for both users and developers. As a way to address this concern, Federated Learning (FL), a Machine Learning model, has been proposed. FL is the process where each client maintains their own local model that learns from their own data, which afterwards is sent over the network to a centralized server to be aggregated into a global model. Despite that, this model, being dependent on user contributions, is not without its flaws when security comes to mind. Prior research has found that threat actors can attempt to harm the accuracy and/or otherwise alter the model’s functionality in a number of ways. This has resulted in creating and improving defensive strategies for FL becoming a field of growing interest. However, most strategies used have only focused on a single type of attack. In this work, we attempt to create Delphi, a “Unified Federation” method that is robust against various threats through the utilization of a Zero-Trust policy and leveraging existing methods in a process known as method chaining.
DelphiFL: An Investigation into a More Robust Federated Learning Model Using Method Chaining and Zero-Trust
Technology has had a significant influence on the lives of today’s generation. However, over the years, protecting private information while maintaining its utility has become a major concern for both users and developers. As a way to address this concern, Federated Learning (FL), a Machine Learning model, has been proposed. FL is the process where each client maintains their own local model that learns from their own data, which afterwards is sent over the network to a centralized server to be aggregated into a global model. Despite that, this model, being dependent on user contributions, is not without its flaws when security comes to mind. Prior research has found that threat actors can attempt to harm the accuracy and/or otherwise alter the model’s functionality in a number of ways. This has resulted in creating and improving defensive strategies for FL becoming a field of growing interest. However, most strategies used have only focused on a single type of attack. In this work, we attempt to create Delphi, a “Unified Federation” method that is robust against various threats through the utilization of a Zero-Trust policy and leveraging existing methods in a process known as method chaining.