Research

I develop methods to understand, audit, and improve machine learning systems. My work spans theory and empirical methods, from training-data memorization and privacy guarantees to evaluation in language, vision, and multimodal applications.

Privacy & memorization · Language model evaluation · Efficient & multimodal learning

Privacy & memorization

What do models remember, and what does that reveal about their training data?

I study the connections between memorization, generalization, and privacy, and develop efficient methods for estimating memorization and auditing data exposure. I also work on differentially private learning and synthetic data.

  • Foundations: theoretical links between differential privacy, memorization, and input loss curvature (ICML 2024).
  • Auditing: black-box membership inference using input loss curvature (NeurIPS 2024, Spotlight).
  • Practical methods: efficient memorization estimates (ICML 2025; ICLR 2026), private image synthesis (TMLR 2024), and mental-health classification (Frontiers in Digital Health 2025).
Explore the papers and methods

I investigate the relationships between memorization, generalization, differential privacy, and the geometry of a model’s loss around individual samples. In Unveiling Privacy, Memorization, and Input Curvature Links (ICML 2024), we establish theoretical connections between these quantities. Curvature Clues (NeurIPS 2024, Spotlight) uses input loss curvature for black-box membership inference, examining whether a sample belonged to a model’s training set.

Our subsequent work develops efficient, theoretically grounded memorization estimates: Towards Memorization Estimation: Fast, Formal and Free (ICML 2025) and Memorization Through the Lens of Sample Gradients (ICLR 2026). These methods help make analysis of training-data influence and memorization practical at scale. I also study out-of-distribution detection, including additive angular margin loss (CVPR Workshops 2025).

Privacy-preserving learning and synthetic data

I develop approaches for learning from sensitive data while controlling what models and released datasets reveal. DP-ImgSyn (TMLR 2024) combines dataset alignment and visual obfuscation with differential privacy for image synthesis.

In DP-CARE (Frontiers in Digital Health 2025), we apply differentially private learning to mental-health classification of social media posts. Together, these projects connect privacy analysis with practical methods for data sharing and model training.

Language model evaluation

How well do models perform across languages and specialized tasks?

I contribute to benchmarks and evaluation studies that test language models in financial and healthcare settings, including Greek-language tasks and expert assessment of medical summaries.

  • Financial NLP: Plutus for Greek finance (EMNLP 2025) and MultiFinBen for multilingual, multimodal evaluation (ACL 2026).
  • Healthcare: multimodal stress detection (BioNLP 2025) and psychiatrists’ evaluation of LLM-generated systematic-review summaries (CL4Health 2026).
Explore the papers and methods

I contribute to benchmarks that test language models across languages, modalities, and specialized tasks. Plutus (EMNLP 2025) evaluates financial language understanding in Greek. MultiFinBen (ACL 2026) extends financial evaluation across languages and text, vision, and audio, with tasks of varying difficulty.

My healthcare NLP work includes multimodal stress detection (BioNLP 2025), combining social media text with synthesized visual information, and expert evaluation of LLM-generated medical evidence summaries (CL4Health @ LREC 2026). The latter examines whether summaries preserve clinically important information and meet psychiatrists’ criteria for professional acceptability.

Efficient & multimodal learning

How can models use less computation while learning from richer signals?

I investigate model compression, hardware-aware learning, and representations that exploit the structure of compressed video. I also study generative augmentation across modalities.

  • Model and hardware efficiency: mixed-precision quantization (IEEE Access 2021) and hybrid RRAM–SRAM computing (DATE 2022).
  • Video understanding: progressive knowledge distillation and unified spatio-temporal representations (WACV 2026).
  • Cross-modal learning: generative augmentation for biological classification (TMLR 2026).
Explore the papers and methods

My work on efficient models spans mixed-precision quantization (IEEE Access 2021), hybrid RRAM–SRAM in-memory computing (DATE 2022), and compressed-video understanding. For video, I investigate progressive knowledge distillation and unified spatio-temporal representations (WACV 2026) that exploit motion vectors, residuals, and intra-frames to reduce computation.

I also work on cross-modal generative augmentation for biological classification (TMLR 2026), exploring how information from one modality can support learning in another. My earlier research on recurrent-network stability (IJCNN 2019) uses eigenvalue spectra to analyze training dynamics.

Full publication list · Curriculum vitae · Discuss research or collaborations

Implementation expertise

I build prototypes and evaluation pipelines in Python and PyTorch, with experience in C/C++, Bash, Linux, Docker, and MATLAB. My work includes mixed-precision quantization with up to 6× network-size reduction on evaluated models, knowledge distillation and early exits for compressed-video recognition, and hardware–software co-design. Results are specific to the experimental settings described in the linked papers.

Code: DP-ImgSyn · Additive angular margin out-of-distribution detection