Privacy-Preserving Artificial Intelligence in Digital Learning: A Systematic Review of Federated Learning, Differential Privacy, Security, and Educational Utility in Higher Education
This systematic review of 33 studies on privacy-preserving AI in higher education finds that while techniques like federated learning and differential privacy can maintain learning utility, the field currently lacks rigorous security evaluations, real-world multi-institutional deployments, and comprehensive assessments of educational impact.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern classroom, learning is no longer just a conversation between a teacher and a student; it is a stream of digital footprints. Every click on a video, every pause on a quiz, and every word typed into a discussion board is recorded. These records hold the power to help educators spot students who are struggling before they fail, to tailor lessons to individual needs, and to plan better courses for the future. However, this same detailed information is deeply personal. It can reveal a student's disability, their home life, their psychological state, or their political views. If this data is gathered into one giant central computer, it becomes a tempting target for hackers and raises serious questions about who gets to see it and how it might be used to judge or limit a person's future.
To solve this, researchers have been developing a set of tools designed to let computers learn from data without ever actually seeing the raw, private records. One approach, called federated learning, is like a group of neighbors who want to learn how to bake a better cake together without ever inviting each other into their kitchens. Instead of sending their secret family recipes to a central baker, they each bake a small batch in their own homes, share only the instructions on how to improve the recipe, and then combine those instructions to create a better master recipe. Another method, known as differential privacy, works by adding a tiny amount of mathematical "noise" or static to the data, similar to how a radio station might broadcast a signal with a slight hiss to prevent anyone from isolating a single voice. This ensures that the final result is useful for the group, but no single person's contribution can be traced back to them. The question facing universities today is whether these clever tricks can actually work in the real world to protect students while still helping them learn.
A team of researchers set out to find the answer by looking at every available study on this topic published between 2018 and September 2026. They sifted through nearly four thousand scientific papers, narrowing their focus to thirty-three studies that tested these privacy tools in higher education settings. Their goal was to see if these technologies had moved beyond the drawing board to become reliable, useful tools for real colleges and universities. What they found was a landscape of rapid technical progress that has not yet caught up with real-world application. The researchers discovered that the technology works well in controlled experiments. In fifteen of the studies, the "neighborhood" approach of federated learning allowed computers to learn from student data without moving the data itself, achieving results almost as good as if all the data had been gathered in one place. In other studies, adding the "static" of differential privacy or creating fake, synthetic student records allowed researchers to build models that could predict things like who might drop out of a course, all while keeping individual identities safe.
However, the study revealed a significant gap between what happens in a computer simulation and what happens in a university. Most of the thirty-three studies were not tests on live students across multiple campuses; they were experiments run on old data or on simulated groups of students created by computers. The researchers found that while the privacy tools often preserved the ability to make accurate predictions, they rarely proved that they could withstand a determined attack. Very few studies tested whether a hacker could still figure out a student's identity by looking at the shared instructions, or whether a malicious actor could trick the system into learning the wrong things. Furthermore, the studies mostly measured success by how accurate the computer's predictions were, rather than by whether the students actually learned more, stayed in school longer, or felt safer. The technology showed it could keep data private in a test tube, but it had not yet been proven to work in the messy, complex reality of a university system where budgets, laws, and human trust matter just as much as code.
The review also highlighted that the field is still in its early stages of maturity. While the number of studies has grown quickly, with more than half appearing in just the last two years, the evidence remains fragmented. Many researchers relied on the same few public datasets, which means the tools have been tested on a narrow slice of educational reality rather than the diverse world of different colleges, languages, and cultures. The researchers noted that in the few cases where privacy was formally measured, the protection was often strong, but in many others, it was merely assumed because the data stayed on a local server. This assumption is dangerous, as the study explained that keeping data local does not automatically stop it from being reconstructed or inferred by others. The authors concluded that while the promise of privacy-preserving artificial intelligence is real, it is not yet a finished product. Universities cannot simply buy a system and expect it to be safe and effective; they must treat these tools as promising infrastructure that requires careful, continuous testing and governance. The path forward, the researchers suggest, involves moving away from isolated computer simulations and toward real-world collaborations between multiple universities, where these systems can be tested against actual attacks, real costs, and the true measure of whether they help students succeed.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.