Recovered Integrated Circuit (IC) netlist analysis is vital for hardware assurance. While powerful open-source tools like MPI-SP’s HAL exist, their adoption is often hindered by the steep learning curve associated with Python programming and tool-specific APIs. Recent advancements in Generative AI offer a solution to bridge this gap.
In this talk, we present an experimental, fully local Retrieval-Augmented Generation (RAG) system designed to automate netlist analysis within HAL. By leveraging Open WebUI and Ollama, we augmented Large Language Models (LLMs) with HAL’s wiki and API documentation. This allows users to instruct the LLM in natural language to generate Python scripts for complex analysis.
We evaluated eight RAG-powered LLMs, comprising six state-of-the-art open models and two commercial models, on tasks ranging from general usage questions to dataflow and HAWKEYE analysis on cryptographic netlists. We verified the correctness of the generated scripts, as well as the response time and memory usage. Our results indicate that reasoning models generally outperform non-reasoning ones, and open local models can achieve parity with leading commercial models. These findings demonstrate that privacy-centric local LLM assistants for hardware security analysis is becoming a viable reality.