Zahra Ghodsi, assistant professor of electrical and computer engineering

Zahra Ghodsi, assistant professor in Purdue University’s Elmore Family School of Electrical and Computer Engineering, has received a U.S. National Science Foundation (NSF) CAREER Award to develop new tools that could make large-scale artificial intelligence more open, reliable, and secure.

The NSF CAREER Award is one of the agency’s most prestigious honors for early-career faculty, supporting researchers who have the potential to serve as academic role models in research and education.

Ghodsi’s project, “Distributed Large-scale Machine Learning with Security Guarantees,” focuses on a growing challenge in artificial intelligence: Today’s most powerful machine learning systems require enormous amounts of data and computing power, which means large-scale AI development is often limited to a small number of organizations with the resources to build and run them.

That concentration can make AI systems less transparent, harder to audit, and more prone to reliability issues when users must rely on only a few service providers. Ghodsi’s research explores allowing many organizations and individuals to contribute data and computing resources within a large-scale machine learning ecosystem while protecting privacy and guarding against malicious actors.

“Large-scale AI development should not be opaque and limited to a small number of players,” Ghodsi said. “This project is about building the foundations for machine learning systems that are more open and distributed, while still providing the security and privacy assurances people need in order to trust them.”

Distributed machine learning could enable organizations, universities, and even individuals to contribute to AI systems in a more transparent way, focusing on applications that support the public good. But a distributed setting also creates new risks. A bad actor could try to manipulate data, interfere with training, or compromise any other part of the process.

Ghodsi’s project aims to address those risks by developing verification methods for distributed and complex machine learning pipelines, including systems that use private data. The work develops new techniques that allow data owners to contribute to the learning process while attesting to the legitimacy of their data contribution and protecting the privacy of any sensitive information.

The project also explores ways to ensure the integrity of machine learning training when it is distributed across a network of potentially untrusted workers. Ghodsi’s team will study methods for proving that computations were performed correctly and will examine how practical factors such as hardware differences and runtime optimizations can influence those outcomes. The research aims to establish security requirements and identify appropriate security parameters that guarantee trustworthy results while maintaining high system efficiency.

In addition to the research, the project includes educational activities designed to help students better understand both the opportunities and vulnerabilities of AI tools, as well as outreach activities that engage stakeholder communities and industry practitioners.