Research examining the critical tradeoff between leveraging proprietary IP for fine-tuning LLMs and preventing IP leakage during inference.

Key insights:

• Fine-tuning LLMs with proprietary data significantly improves performance for niche languages like Verilog
• However, this creates substantial IP leakage risks through model inference
• The research quantifies this security-utility tradeoff for hardware design companies
• Offers practical strategies for balancing model utility and IP protection

This work addresses a fundamental challenge for engineering firms seeking to deploy AI coding assistants while safeguarding their intellectual property, with implications for how companies approach LLM customization in specialized technical domains.

VeriLeaky: Navigating IP Protection vs Utility in Fine-Tuning for LLM-Driven Verilog Coding