Marc Molina
Van den Bosch
Privacy Engineer & Machine Learning Researcher
Industrial PhD @ CERN × Universitat Pompeu Fabra · Geneva, Switzerland
I work on the part of machine learning where the guarantee matters as much as the accuracy: differential privacy, secure aggregation, and federated learning, composed into distributed-DP systems with end-to-end guarantees — plus second-order optimisation that recovers the utility DP-SGD gives away.
Alongside the theory, I ship: CAFEIN, CERN's federated learning platform; a synthetic-data pipeline hospitals actually share at Kantonsspital Aarau; and Kosmico, a collaborative AI-native research platform.
The Missing Sampler: distributed privacy accounting for async FL
Buffered async FL lets fast clients accumulate up to 97× the reported privacy loss. A privatiser committee restores per-client accounting and subsampling amplification.
Distributed DP · Secure aggregation→ ICML 2026DP-KFC: data-free preconditioning for private deep learning
A network's optimisation geometry is architectural, not data-borne: probing with structured synthetic noise builds KFAC preconditioners that beat DP-SGD at ε ≤ 3 — no private or public data.
KFAC · Mean-field theory · Pink noise→ Applied · deployedSynthetic data hospitals can actually share
Multimodally-conditioned generation for cross-hospital sharing, evaluated on quality, diversity, and memorisation.
Diffusion · GANs · Leakage audit→Open to conversations about differential privacy, federated systems, and building things that protect the people in the data.
Kosmico
A collaborative, AI-native research platform — built for how research actually gets done.
kosmico.ai ↗Industrial PhD Researcher, Private & Federated ML
Apr 2025 — nowCERN × Universitat Pompeu Fabra · Geneva
Second-order methods for sample- and user-level DP. Distributed DP combining differential privacy, secure aggregation, and federated learning. CAFEIN platform.
Machine Learning Researcher
Feb 2024 — nowKantonsspital Aarau · Aarau
Privacy-preserving synthetic image pipeline with multimodal conditioning for cross-hospital sharing; evaluation spanning quality, diversity, and memorisation.
Data Science Intern
Aug 2023 — Jan 2024Roche · Basel / Kaiseraugst
Anomaly detection for large-scale clinical-trial wearable and observational data; LLM + RAG system on Azure for trial findings reporting.
Biomedical Computer Vision Developer
Jun 2021 — Aug 2023Vall d'Hebron Research Institute · Barcelona
Graph U-Net on vessel graphs; cerebral artery and large-vessel-occlusion segmentation on CTA volumes for stroke workflows.
The Missing Sampler: Distributed Privacy Accounting for Asynchronous Federated Learning
Under review · PDF ↓
Architecture-Driven Preconditioning for Sample-Efficient Differentially Private Medical Image Segmentation
Under review
The Interplay Between Explainability and Differential Privacy in Federated Healthcare
DECAF
Decoder-Free Supervoxel GNN for Accurate Brain-Tumor Localization in Multi-Modal MRI
GRAIL
Memorization Detection Benchmark for Generative Image Models
Safe Generative AI Workshop
Enhancing federated learning in multicenter studies via synthetic data generation for improved downstream performance
A multifaceted evaluation framework of generative medical image models
Best MSc thesis — Novartis Industry Diplompreis
Conditional Style-based GANs for Generating Synthetically Private Mammograms
Top-5 abstract — KSA Tag der Innovation & Forschung
Evolutionary, Bayesian, and Quasi-Monte Carlo Hyperparameter Tuning
Fully automated Large Vessel Occlusion localization and segmentation from CT and CTA