Miroslav Vacura presented his research at conference Interdisciplinary Information Management Talks held in Hradec Králové on 4 – 6 September 2024, organized by Johannes Kepler University Linz and Prague University of Economics and Business. He presented paper “Watermark as a Tool to Address Abuse of Large-Scale Language Models”.
Presented talk abstract: This paper deals with the problem of identifying texts generated by large language models (LLM) using digital watermarking technology. Advances in artificial neural networks, which are now capable of producing texts comparable to human written speech, have led to an increased risk of misuse of these technologies for ethically questionable purposes such as spreading misinformation or generating academic papers. This paper focuses on the development of methods to incorporate digital watermarks into generated texts, thus enabling their subsequent automatic identification. A suitable watermark should, among other things, be resistant to text editing and easily detectable by special software, while at the same time not being easily removable. Watermarks can take various forms, from the simple insertion of metadata to the use of complex cryptographic methods. This study provides an overview of existing methods, while presenting a new approach that could improve detection capability while keeping computational resource requirements to a minimum.
Leave a Reply