Watermarking LLMs — Challenges and Opportunities in Electronic Design Automation

The recent development of large language model (LLM) has led to their mass adoption on applications across various fields, such as text analysis and generation. LLMs also show their ability to generate effective code from natural language prompts. This has resulted in the transformation of hardware development, including generating integrated circuits (ICs) in codes at the register-transfer level (RTL) using LLMs. However, the proliferation of LLM-generated texts and codes has increased the risk of IP piracy, malicious code generation, and security vulnerabilities, necessitating the need to identify LLM-generated hardware designs.One solution is LLM watermarking, in which identifiable patterns can be embedded into the generated text. Although prior works have identified LLM-generated text and software programming, LLM watermarking for hardware design is largely unexplored. In this paper, we evaluate prior works on watermarks for texts, code, and hardware designs tasks. The emerging LLM-based watermark application strategies are also detailed, in which the opportunities and challenges associated with applying watermarks to LLM-generated hardware codes are defined.