An Advanced Detection Framework for Embedded System Vulnerabilities

Abstract

Embedded systems serve as the foundation for modern computing in industrial, IoT, and defense applications. However, their increased adoption exposes them to security threats across multiple levels, from application software to operating systems and hardware. Traditional security mechanisms often focus on a single layer, missing cross-layer vulnerabilities that can be exploited by sophisticated attacks. This paper introduces a unified multi-layer vulnerability detection framework that integrates software, operating system, and hardware-level security analysis into a single LLAMA3-based deep learning model. The framework leverages a combined dataset consisting of source code vulnerabilities (C programming), Linux Kernel exploits (system calls), and Field-Programmable Gate Array (FPGA) hardware security risks in Hardware Description Languages (HDLs) such as Verilog and VHSIC Hardware Description Language (VHDL) and bitstream analysis. By merging these diverse vulnerability types into a single learning model, the framework is capable of detecting security threats across the entire embedded system stack. Applying static and dynamic analysis techniques across multiple layers, the proposed LLAMA3-based model achieves state-of-the-art detection accuracy across embedded system security domains. Experimental results demonstrate that the integrated framework outperforms existing layer-specific models, achieving 99.30% accuracy for software vulnerabilities, 99.16% accuracy for OS-level exploits, and 95.6% accuracy for hardware security threats.