A Systems Theory Approach to Cybersecuring a Supervised Machine Learning System

Abstract

Machine learning is a rapidly growing field with many applications in areas such as healthcare, finance, and transportation. As machine learning becomes more prevalent, it is important to ensure that these systems are secure and can resist attacks from malicious actors. This is particularly difficult because Machine Learning has become a black box, meaning that the models used to perform machine learning tasks can be very complex and might include millions or billions of parameters. This complexity makes it difficult to understand how the model makes decisions or predictions, and it can be hard to explain why the model produced a particular output. It is here where a systems approach can be helpful since it can understand and analyze complex systems and their interactions as a whole. It involves considering the relationships and interactions between the parts of a system, rather than just the individual parts themselves.

This thesis aims to adopt a systems approach to security in machine learning systems using System-Theoretic Process Analysis for Security (STPA-Sec). Due to the broadness of the field, this thesis focuses on Supervised Machine Learning Systems and provides generalized recommendations.