UCF news

 https://www.cs.ucf.edu/ucf-researcher-aims-to-advance-ai-security-with-nsf-career-award/

tificial intelligence (AI) has become a valuable tool for data analysis across the finance, healthcare and government sectors. But AI has a significant flaw – it can expose private and sensitive data during computation.

Professor Qian Lou of the UCF Department of Computer Science plans to change the way data is protected by AI with the support of a $600,000 grant from the National Science Foundation’s Faculty Early Career Development (CAREER) Program, which supports early-career faculty who have the potential to serve as role models in research and education.

Lou’s objective is to enable AI to perform computations on protected data without exposing the underlying information to the cloud provider or to unauthorized users.

“My goal is to build security into the AI infrastructure itself—from secure algorithms and GPU kernels to compilers and software libraries,” Lou says. “If encrypted computation becomes fast, scalable and easy to use, organizations can benefit from AI without surrendering the privacy of their data or models.”

How AI Queries Work

Medical facilities could use AI to analyze medical images or clinical records. Banks might employ AI models to detect fraud and money laundering schemes. National security agencies may use AI for defense purposes. In each of these instances, the use of sensitive data would be required, but that data would no longer be protected.

When an AI query draws information from private data, the data is encrypted while it is sent to the server. As the data enters the AI system or the graphics processing unit (GPU) memory, it is decrypted for inference.

“During that time, raw inputs, model activations, and sometimes proprietary model parameters may be exposed to the cloud operator, a malicious insider, or an attacker who compromises the host,” Lou says. “Our project seeks to keep sensitive values protected while computation is running.”

The Challenge of AI Security

As Lou and his team develop their research, there are several challenges to overcome. The first is performance. While modern GPUs are optimized for basic AI computation, they can’t adequately handle encrypted mathematics. This means that security compromises speed and efficiency.

AI also involves many different operations include nonlinear functions, inferences and training models. Different security measures may be needed for each function, so there isn’t a one-size-fits-all security method that will work well for every task.

Another obstacle is usability. Secure AI requires deep expertise in cryptography, which not every developer had. Lou’s goal is for encryption to be automated so that developers can focus on writing programs while the system secures itself.

A Novel Framework for Secure AI

The idea behind Lou’s AI framework is simple: embed security into the platform as it is built.

“Our framework treats security as an AI-infrastructure problem,” Lou says. “It co-designs the mathematics, cryptography, GPU and accelerator kernels, and software automation. A key idea is algebraic invariance—reorganizing encrypted computation to avoid unnecessary ciphertext transformations. The project also develops integer and modular kernels for common accelerators and uses hybrid protection when that is the better fit. The goal is a practical stack, not a single algorithm.”

The end result will be open-source software that eliminates the need for specialized cryptography and connects to existing AI frameworks. Additionally, students will also gain hands-on experience in AI, cybersecurity, GPUs and applied cryptography.

“The project will lead to the creation of course and laboratory modules, research opportunities, mentoring, public tutorials and K-12 outreach,” Lou says. “Students will learn not only how AI models work, but also how the underlying infrastructure can be redesigned for security.”