Collaborative Research: CISE Crosscutting Small: SaTC: SAGE: Secure Accelerators for Next-Generation Foundation Models

Howard University, Washington DC

Investigators

Abstract

Foundation models are general-purpose technologies that power a wide range of artificial intelligence (AI) applications, including intelligent chatbots, voice assistants, cyber threat detection systems, and autonomous robots. These models can also be adapted to new tasks via fine-tuning techniques. However, the local deployment of foundation models on consumer devices is challenging due to the models’ large size and high computational demands. The models also face significant security risks, such as intellectual property (IP) theft and malicious tampering, when deployed outside of secure platforms. To address these challenges, this project will build modular hardware accelerators that enable secure and efficient deployment of fine-tuned foundation models in consumer devices. These accelerators can be securely integrated into existing AI hardware systems and will play a critical role in enhancing the security and resilience of the United States AI semiconductor supply chain. As part of the design process, the team will first build a heterogeneous system containing a graphics processing unit (GPU) and a custom accelerator. The GPU will store the open parameters of a foundation model, and the accelerator will support the secure execution of the fine-tuned component. This prototype will be used to evaluate performance bottlenecks and security vulnerabilities of this heterogeneous system. Based on the findings, the team will then devise advanced acceleration methodologies and implement active locking mechanisms to protect the fine-tuned model. The final phase of the project will produce a secure and modular accelerator realized on a physical platform. Overall, this project will address the challenges of deploying foundation models in resource-constrained devices, deliver secure hardware designs to safeguard foundation models, and advance interdisciplinary research between AI acceleration and hardware security. The solutions developed will also be disseminated to the public via open-source platforms, encouraging the broader and industrial adoption of the products. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.