Analyzing AI Image Recognition Utilizing Low-Power GPU Architecture in Deterministic Safety-Critical Multi-Core Avionics Systems

This paper presents a novel approach to utilizing image recognition on low-power GPU systems, specifically targeting airworthiness analysis within defense applications. Unlike traditional multi-core processors, GPUs operate as massively parallel many-core systems, presenting unique challenges for the integration of AI models within software airworthiness. Our research aims to expand upon established analysis tests and metric data on the NVIDIA Orin AGX module to investigate interference channels in AI GPU systems. The experiment involves running controlled image recognition tasks while introducing adversarial applications to analyze potential interference and system performance under varying conditions. This systematic approach will evaluate critical hardware metrics, including power efficiency, and provide a comprehensive characterization of the GPU's capabilities. The findings will contribute to a clearer understanding of the feasibility and pathways toward certification and airworthiness standards for AI-integrated systems in military aviation, ultimately aiding in the development of reliable and efficient AI methodologies for defense applications.