Abstract

In the ever-evolving landscape of artificial intelligence (AI), the vulnerability of AI models to adversarial attacks has become a critical concern. The problem at hand lies in the lack of dedicated security mechanisms tailored for FPGA-based AI implementations, leaving them exposed to threats such as tampering, reverse engineering, and unauthorized access. This research addresses the pressing need for robust security measures by proposing a hardware-based approach to protect FPGA (Field-Programmable Gate Array) implementations of AI models. FPGAs offer a flexible and efficient platform for deploying AI models but are susceptible to attacks that compromise the integrity of the implemented algorithms. This involves the integration of specialized security modules within the FPGA architecture. These modules are designed to detect and thwart various forms of attacks, including side-channel attacks and unauthorized access attempts. Leveraging the inherent parallelism and reconfigurability of FPGAs, the security modules operate seamlessly alongside the AI model, imposing minimal overhead on performance. Results from experimental evaluations demonstrate the effectiveness of the hardware-based security approach in preventing unauthorized access and tampering with the FPGA-based AI model. The proposed solution showcases resilience against common attack vectors, ensuring the confidentiality and integrity of the deployed AI models.