Skip to main content

NMSU alumnus gains international recognition for research on privacy and fairness in AI

Original article:

Ferdinando Fioretto, who received dual doctorates in computer science from New Mexico State University and the University of Udine in Italy, has won multiple awards in the past year for his influential research on privacy and fairness in machine learning algorithms as well as on the integration of operations research with deep learning.

For his pivotal contributions, Fioretto, an assistant professor of computer science at Syracuse University, has been awarded an Early Career Researcher Award by the Association for Constraint Programming, a Young Investigator Award for Research in Computer Science by the Italian Scientists and Scholars in North America Foundation, a National Science Foundation CAREER Award, and a Google Research Scholar Award.

In addition to his NSF and Early Career Awards, Fioretto has earned multiple Best Paper Awards, including a Best AI Dissertation Award by the International Conference of the Italian Association for Artificial Intelligence.

“As we evolve as a society,” Fioretto said, “algorithms are replacing some of the decisions that were made by us humans.” These algorithms are used in various domains of our society, such as federal fund allocation, energy systems, criminal justice, loan approvals, and ranking systems like those used in employment decisions.  

Algorithms are fed historical data that they analyze to build mathematical models that can be used on subsequent data to make predictions or “abstractly reason to give us an outcome,” Fioretto explained. In this process, termed “machine learning,” algorithms frequently introduce or employ existing biases, which in turn seep into the decisions they make. Consequently, these decisions can come with huge societal impacts.

For example, what’s known as the PATTERN algorithm is used to predict whether a prisoner will be likely to reoffend based on data from previous court cases, which were decided by people with their own, even unconscious, biases. Unfortunately, the algorithm learns these biases as it creates a predictive model from the data.

These biases can arise from latent factors, variables that aren’t directly observed but can be inferred, which are “interpreted or misinterpreted as relevant,” Fioretto said. These factors can be associated with individuals’ protected attributes, such as race and gender.

In 2006 there was a “breakthrough,” as Fioretto described it, in the field of theoretical computer science called “differential privacy,” which is a mathematical way to ensure that data generated by an algorithm cannot be used to infer any identifiable information about the individuals who are described in the data. Among others, the US Census Bureau is adopting this privacy definition in its 2020 data release.

“This technology relies on noise addition to ensure that the algorithm’s outputs remain insensitive to the participation of any individual,” Fioretto explained. However, this noise affects the accuracy of the algorithms’ outputs. “Once you're adding noise, you're altering some properties of the data,” he said. “My group has found that the mechanisms that are adopted in order to satisfy privacy properties can make errors that are actually disproportionate with respect to some characteristics of the population and exacerbate the existing bias toward minorities”.

In other words, implementing privacy measures on data used for decision processes can have significant societal and economic repercussions.

Fioretto’s research looks particularly at how allocation decisions are affected. Title I, the program that funds low-income school districts, uses rule-based algorithms to decide where and how much money is distributed. He explained that depending on the privacy preservation of the data used, school districts might receive substantially less money than what would have been warranted.  

Fioretto added that “the relationship between privacy, accuracy and fairness is complicated.” His research is working towards disentangling this relationship to devise algorithms that can guarantee the privacy of individuals without excessively sacrificing the accuracy of the data and the fairness of the decisions made.

Fioretto not only searches for solutions to the large-scale consequences of machine learning algorithms, but also engages the community around the importance of these issues. He co-organizes the annual AAAI workshop on Privacy Preserving Artificial Intelligence and works with policymakers and data users to understand and address these concerns. Fioretto’s research and activism will ultimately bring us closer to more equitable decisions being made for our communities.