Analysis of Watermarking Methods in Large Language Models

This paper is devoted to the analysis of methods of Large Language Models (LLMs) protection through the investigation and application of various post-hoc and watermarking techniques, including WLLM (A Watermark for Large Language Model), SWEET (Selective WatErmarking via Entropy Thresholding) and three newly proposed watermarking methods. The analysis is divided into three parts. Firstly, we investigate relevant literature on previous research on LLMs protection algorithms. Then, the analysis aims at subsequently scrutinizing both theoretical and empirical directions of the research topic, where the former one is designed to inspect and understand the fundamental aspects of LLM-generate text detection methods, and the latter’s purpose is to observe the performance rates of some of the covered watermarking techniques in real-world settings. The issue of preventing LLM-generated text from unauthorized usage and distribution has recently gained acknowledgment. A lot of different mathematical models were already introduced. These models may be largely divided into the post-hoc and watermarking streams, while we find out that the latter one is more efficient and robust to adversarial influence. Though the research theme is quite topical, almost none of the existing papers cover protection of LLM-generated programming code – unrecognized, though extremely purposeful topic. The distinguishing feature of our work is to scrutinize methods of generating and detecting watermarks in programming code, which could allow to popularize the corresponding topic, give future researchers basis for the analysis and apply introduced LLM watermarking methods in practice.