Gary Gary KLAJER

I conducted a security audit of Medinote, a medical language model, focusing on patient data leakage risks. My research introduced a novel approach by executing targeted extraction attacks using auxiliary public data to recover private information. I explored how statistical patterns, such as distributions of patient attributes, were encoded within the model and developed sequence-based perplexity metrics to distinguish in-distribution from out-of-distribution data. This could enhance the filtering of non-member data from generated candidate samples.

This work was done at PariSanté Campus within the MILES team, under the supervision of Prof. Jamal Atif and Prof. Olivier Capet and collaboration of Danil Savine. The MILES project-team is hosted by the LAMSADE research unit at Université Paris Dauphine-PSL, France and is heavily involved in the Paris Artificial Research Intelligence Institute (PRAIRIE), a national AI research initiative funded by the 3AI Plan.

In addition to my research work, I:
• Conducted a comprehensive literature review of over 100 papers covering topics such as medical LLMs, memorization, extraction attacks, membership inference attacks, DP-preserving techniques, unlearning, in-context learning (ICL), and cutting-edge transformer architectures (e.g., KNN-LLM, Matryoshka Embedding, Mixture of Experts, State Space Models).
• Re-implemented bigram models, BPE, decoder-only transformers, and MoE from scratch using PyTorch to deepen my understanding of these architectures.
• Re-proved theorems on benign overfitting from Bartlett et al.'s work.
• Fast-tracked Differential Privacy (DP) by following the "DP for ML" course from M2 IASD @ Paris Dauphine and completing Dwork's textbook and programming notebooks on DP.
• Attended reading groups on advanced mathematical topics like higher-order optimization, optimal transport, automatic architecture search, etc.