Trustworthy & efficient machine learning
Efstathia Soufleri
Postdoctoral Researcher · Archimedes Unit, Athena Research Center
I study what machine learning models memorize, how to protect sensitive data, and how to evaluate models beyond benchmark scores.
My work connects theoretical analysis with practical methods for privacy auditing, language-model evaluation, and efficient multimodal learning. Ph.D. in Electrical and Computer Engineering, Purdue University.
Research at a glance
Privacy & memorization
Understand training-data exposure and develop efficient methods for memorization estimation and privacy auditing.
Explore this work →Language model evaluation
Evaluate models across languages, modalities, and specialized financial and healthcare tasks.
Explore this work →Efficient & multimodal learning
Reduce computation through quantization and compressed-video representations, and learn across modalities.
Explore this work →Selected contributions
NeurIPS 2024 · Spotlight
Curvature Clues
Problem: audit whether models expose their training data.
My contribution: co-developed curvature-based privacy attacks and analysis connecting loss geometry with memorization.
Outcome: black-box membership inference; NeurIPS Spotlight.
ICLR 2026
Memorization Through the Lens of Sample Gradients
Problem: memorization estimates can be expensive to compute.
My contribution: developed sample-gradient-based proxies and fast, formal estimators.
Outcome: scalable methods reported at ICLR 2026 and ICML 2025.
ACL 2026
MultiFinBen
Problem: standard benchmarks miss multilingual financial tasks.
My contribution: built evaluation pipelines for Greek finance and multilingual, multimodal benchmarks.
Outcome: Plutus (EMNLP 2025) and MultiFinBen (ACL 2026).
TMLR 2024
DP-ImgSyn
Problem: share useful image data while protecting privacy.
My contribution: designed a discriminative synthesis framework with formal differential privacy guarantees.
Outcome: evaluated synthetic-data utility through downstream classification.
View all publications · Research overview
Methods & implementation
I develop research prototypes and training and evaluation pipelines in Python and PyTorch. My expertise includes differential privacy, membership inference, memorization estimation, benchmark development, model compression, and knowledge distillation.
Recognition
ICML 2026 Gold Reviewer · NeurIPS 2025 Top Reviewer · NeurIPS 2024 Spotlight
Research opportunities
I am interested in research positions in academia and industry focused on trustworthy machine learning, privacy, model evaluation, and efficient learning. Get in touch to discuss opportunities and collaborations.