Mapping LLM Training Data for Safety and Explainability: Approach of Low Perplexity
As Large Language Models (LLMs) become embedded in products and services, their reliance on vast, often opaque training data raises pressing risks around safety, intellectual property, and trust. A central question for privacy, compliance, and safe deployment is: when an LLM is reproducing memorized sequences versus generalizing?
About the speaker
Dr. Anastasiia Kucherenko
Scientific Collaborator
at
HES-SO Valais-Wallis (IEM)
Dr. Anastasiia Kucherenko is a postdoctoral researcher at HES-SO Valais-Wallis, Switzerland, working at the intersection of AI safety and cybersecurity in collaboration with the Cyber Defense Campus.
Her research focuses on the safety of large language models, developing methods to trace and evaluate their training data and prevent harmful or biased outputs.
She completed a PhD in Computer Science at EPFL on robustness and anonymity in large-scale distributed systems, and held cryptography research internships at Microsoft Research and the Institute of Science and Technology Austria.
Her work has been presented at top conferences including ACL, CCS, DISC, and SRDS.
Read more …
Read more …