Together with the Fraunhofer Institute for Digital Media Technology IDMT in Ilmenau and the Artefact Germany in Hamburg, scientists from the Faculty of Media are researching digital watermarks texts and spoken language. Their goal is to make automatically generated content identifiable, thus increasing transparency and trustworthiness in digital communication spaces. Users should be able to determine which programme was used to create a text or whether a text was written or spoken by a human.

The research project is focussing on watermarking, a process that allows markings to be hidden as part of a text with the aim of identifying the origin of the text or recognising text modifications. An additional goal of the project is to modify large language models (LLMs) in order to »brand« them. In other words, the models would bear a sort of digital signature that can be used to distinguish branded language models from non-branded ones. Companies or individuals can use these certified language models to detect unauthorised statements about themselves by identifying those that are not watermarked.

Lantmark’s combination of text and spoken analysis is currently something unique: The researchers are proposing an innovative procedure for bimodal authentication that checks written and spoken content together. This could be used, for instance, to identify statements that have been »put in someone’s mouth« using speech synthesis.

One of the project’s goals is to develop methods that can be tangibly used, for instance to protect against digital manipulation. »Large language models like ChatGPT are being used in numerous professional and private settings. Digital watermarks and ›certified‹ language models have the potential to provide proof of origin in the future«, explains Prof. Dr. Benno Stein, project coordinator and Director of the Professorship of Intelligent Information Systems at the Bauhaus-Universität Weimar.

»Lantmark – Potential und Grenzen von Large Language Models für Anwendungen im Watermarking von Text und gesprochener Sprache« is a joint project that is funded by the Federal Ministry of Education and Research as part of the »Sichere Zukunftstechnologien in einer hypervernetzten Welt: Künstliche Intelligenz« initiative with nearly 1.97 million euros.

Additional information on the research project can be found at: https://www.forschung-it-sicherheit-kommunikationssysteme.de/projekte/lantmark 

In case of questions, please contact Tina Meinhardt, Research Communication and Public Relations Representative at the Faculty of Media, by telephone at +49 (0) 36 43 / 58 37 65, or by e-mail at tina.meinhardt[at]uni-weimar.de.

Together with the Fraunhofer Institute for Digital Media Technology IDMT in Ilmenau and the Artefact Germany in Hamburg, scientists from the Faculty of Media are researching digital watermarks texts and spoken language. Their goal is to make automatically generated content identifiable, thus increasing transparency and trustworthiness in digital communication spaces. Users should be able to determine which programme was used to create a text or whether a text was written or spoken by a human.

The research project is focussing on watermarking, a process that allows markings to be hidden as part of a text with the aim of identifying the origin of the text or recognising text modifications. An additional goal of the project is to modify large language models (LLMs) in order to »brand« them. In other words, the models would bear a sort of digital signature that can be used to distinguish branded language models from non-branded ones. Companies or individuals can use these certified language models to detect unauthorised statements about themselves by identifying those that are not watermarked.

Lantmark’s combination of text and spoken analysis is currently something unique: The researchers are proposing an innovative procedure for bimodal authentication that checks written and spoken content together. This could be used, for instance, to identify statements that have been »put in someone’s mouth« using speech synthesis.

One of the project’s goals is to develop methods that can be tangibly used, for instance to protect against digital manipulation. »Large language models like ChatGPT are being used in numerous professional and private settings. Digital watermarks and ›certified‹ language models have the potential to provide proof of origin in the future«, explains Prof. Dr. Benno Stein, project coordinator and Director of the Professorship of Intelligent Information Systems at the Bauhaus-Universität Weimar.

»Lantmark – Potential und Grenzen von Large Language Models für Anwendungen im Watermarking von Text und gesprochener Sprache« is a joint project that is funded by the Federal Ministry of Education and Research as part of the »Sichere Zukunftstechnologien in einer hypervernetzten Welt: Künstliche Intelligenz« initiative with nearly 1.97 million euros.

Additional information on the research project can be found at: https://www.forschung-it-sicherheit-kommunikationssysteme.de/projekte/lantmark 

In case of questions, please contact Tina Meinhardt, Research Communication and Public Relations Representative at the Faculty of Media, by telephone at +49 (0) 36 43 / 58 37 65, or by e-mail at tina.meinhardt[at]uni-weimar.de.