SaTC: CORE: Small: Developing Holistic Systems to Secure the Machine Learning Supply Chain

Purdue University, West Lafayette IN

Investigators

Abstract

Machine learning (ML) systems have made impressive strides in recent years, creating complex models that are able to perform complicated tasks from processing language to generating realistic images and videos. In order to create such systems, we require a similarly complex network of actors and activities, including data collection and cleaning, the use of existing ML providers and libraries, and validation and post-deployment testing. Much like supply chains in the context of open source software, the complexity of ML networks allows for rapid and effective development of new ML systems. But also like open source software, this supply chain creates risks: a malicious actor who can attack any step of the process may be able to impact the safety, quality, and privacy of ML models, and of the people and organizations who use them. This project aims to create holistic systems to define and quantify trust in machine learning supply chains and develop tools to verify and audit specifications and requirements across those supply chains. The research will also support the development of a diverse next generation of computer scientists by incorporating outcomes in coursework and research opportunities. The project is divided into three main tasks. The first task investigates different pillars of trustworthy machine learning (such as privacy, robustness, or fairness) as a whole. The main goal for this task is to empirically understand the interactions between different pillars and their trade-offs. The second task constructs system auditing mechanisms to verify the security of the machine learning supply chain as described by trustworthy machine learning requirements identified in the previous task. Finally, the third task creates a comprehensive and universal framework as the scaffolding for other researchers and developers to integrate, test, validate, and evaluate new mitigations into the framework. In particular, the tools developed in this task will allow for quantitative measures of the effect of new techniques on all pillars of trustworthy machine learning, and the system as a whole. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.