Hybrid Structural and Semantic Analysis for RTL-Level Hardware Trojan Detection Using Large Language Models

Abstract

Hardware Trojans pose a significant threat to the security and reliability of modern integrated circuits, as they can enable information leakage, performance degradation, and functional failure. Existing Trojan detection methodologies are primarily based on structural or statistical analysis and often struggle to detect the semantic intent having malicious circuit modifications. In our paper, we propose a large language model (LLM) guided hardware Trojan detection framework that integrates RTL level structural analysis with semantic reasoning. The proposed ap-proach studies Register Transfer Level (RTL) designs from Trust-Hub benchmarks using automated parsing and synthesis tools such as Yosys. Structural characteristics extracted from RTL designs are converted into human readable textual summaries, which are examined by a locally deployed, open-source LLM (Qwen2.5-Coder-7B-Instruct) behaving as a semantic security auditor. The semantic suspicion score constructed by the LLM is fused with structural deviation metrics applying a weighted strategy to give a final detection decision. The proposed framework works without any model training or fine-tuning and can be operated entirely on local systems. While currently evaluated on Trust-Hub AES benchmarks, the framework demonstrates the feasibility of combining semantic reasoning with RTL-level structural analysis for interpretable Trojan assessment. These results highlight the potential of LLM based semantic reasoning as practical and reliable tools in hardware security auditing workflows.