We ensure that AI makes fair, clear, and safe decisions that people can rely on and trust.
Our home is the CISPA Helmholtz Center for Information Security in Saarbrücken, Germany.
We develop mitigation strategies to address the misuse of AI-generated media across various modalities, including images, text, and audio.
Our research focuses on evaluating the effectiveness of fake audio detection systems, particularly in the context of malicious audio used in scams. In parallel, we assess the robustness of AI-generated image detectors, examining their ability to prevent the spread of misinformation and withstand adversarial manipulation.
Our research on large language model (LLM) security explores vulnerabilities such as prompt injection attacks across various application settings, as well as the models’ capacity to maintain confidentiality of contextual information.
Additionally, we investigate the secure use of LLMs for code generation and their potential as reliable tools for security professionals and software developers.
Beyond LLMs, our work encompasses broader security-related topics in machine learning across multiple modalities.
This includes evaluating the security of audio and speech recognition models, as well as exploring the application of continual learning and out-of-distribution detection techniques in security-sensitive contexts.
