Project Grant 2301940

This two-year National Science Foundation project grant of $137,683 aims to enhance the reliability and privacy of graphics processing unit (GPU)-based deep learning computing. Funded under the Computer and Information Science and Engineering program, the award supports research at West Virginia University to address vulnerabilities in GPU architecture and design lightweight protection schemes. Specifically, the university researchers will explore vulnerabilities impacting GPU-based deep learning and evaluate possible solutions at the compute unit level through scheduling algorithms and activation acceleration. The project will also examine selective integrity mechanisms for secure data transfer between the CPU and GPU with minimal performance overhead. Findings will be integrated into courses and outreach activities to benefit education. The goal is to enable deep learning developers and consumers to focus on advancement and use of models without privacy concerns related to the GPU computing platform.