Research

Genomics and rare diseases

Genomic data reveal information about individuals and their relatives, which makes sharing them across sites especially sensitive. We design privacy-preserving methods for genome-wide association studies that scale with cohort size and for querying whole-genome variant databases, and we support rare disease research across institutions: identifying disease-causing variants, analyzing rare disease variants under homomorphic encryption and detecting disease-associated cell subsets in single-cell data.

Key papers

• arXiv preprint 2026 Privacy-Preserving Detection of Rare Disease-Associated Cell Subsets via Secure Multi-Party Computation
• Nature Communications 2025 PP-GWAS: Privacy Preserving Multi-Site Genome-wide Association Studies
• Bioinformatics 2025 PRISM: Privacy-preserving Rare Disease Analysis using Fully Homomorphic Encryption
• Bioinformatics 2022 Efficient privacy-preserving whole-genome variant queries
• Bioinformatics 2020 Identifying disease-causing mutations with privacy protection
• Balkan Journal of Electrical and Computer Engineering 2019 An Active Genomic Data Recovery Attack

Medical record linkage

Combining patient data from different sources can uncover new relationships between diseases and medical indications. We accelerate probabilistic privacy-preserving record linkage (PPRL) with three-party computation, so that records of the same patient can be matched efficiently and securely across institutions.

Key papers

• Journal of Biomedical Informatics 2025 Accelerating Probabilistic Privacy-Preserving Medical Record Linkage: A Three-Party MPC Approach

Secure computation frameworks

We build frameworks for computing on data that no single party can see. CECILIA, our three-party computation framework, extends privacy-preserving machine learning beyond convolutional neural networks to recurrent kernel networks (RKNs), long short-term memory networks (LSTMs) and generative adversarial networks (GANs). We also develop secret-sharing protocols for training models such as logistic regression and support vector machines and for imputing missing medical data, and randomized-encoding approaches that make kernel computations on distributed medical and image data fast and scalable.

Key papers

• IEEE Access 2025 Secure and Efficient Logistic Regression with Secret-Sharing MPC and Differential Privacy
• Patterns 2024 A privacy-preserving approach for cloud-based protein fold recognition
• CIBB 2024 Private, Efficient and Scalable Kernel Learning for Medical Image Analysis
• IEEE HealthCom 2024 Privacy Preserving Data Imputation via Multi-party Computation for Medical Applications
• arXiv preprint 2022 CECILIA: Comprehensive Secure Machine Learning Framework
• AAAI 2021 Escaped: Efficient secure and private dot product framework for kernel-based machine learning algorithms with applications in healthcare

Collaborative model evaluation

We develop methods that let clients evaluate machine learning models on pooled test data without compromising privacy, for example by computing ROC and precision-recall curves securely. This matters for clients with limited data who need to assess models trained on aggregated datasets, and it supports collaborative settings such as federated learning and the PHT-meDIC platform for distributed medical analytics.

Key papers

• IEEE TPS-ISA 2025 Enabling Privacy-preserving Model Evaluation in Federated Learning via Fully Homomorphic Encryption
• PLOS Digital Health 2025 Privacy-preserving AUC computation in distributed machine learning with PHT-meDIC
• NSS 2023 ppAURORA: Privacy Preserving Area Under Receiver Operating Characteristic and Precision-Recall Curves with Secure 3-Party Computation

Hardware-based security

We design security protocols that build on hardware: physically unclonable functions (PUFs) for second-factor authentication and scalable RFID authentication, and trusted execution environments for a privacy-preserving smart grid.

Key papers

• IEEE Access 2023 A Privacy-Preserving Scheme for Smart Grid Using Trusted Execution Environment
• PLOS ONE 2023 P/Key: PUF Based Second Factor Authentication
• Computer Networks 2023 Scalable RFID Authentication Protocol Based on Physically Unclonable Functions

Federated learning

We address the privacy challenges of federated learning with high-dimensional medical data and develop frameworks that keep model updates and data protected throughout collaborative training. This includes federated domain adaptation for sites whose data follow different distributions, and applications such as epileptic seizure prediction from EEG data, diagnosis of rare genetic syndromes from facial images, age prediction from DNA methylation data and the collective defense of IoT networks.

Key papers

• ICML 2026 Unsupervised Multi-Source Federated Domain Adaptation under Domain Diversity through Group-Wise Discrepancy Minimization
• Studies in Big Data (Springer) 2026 A Blockchain-Based Explainable Federated Learning System for the Trustworthy Collective Defense of IoT Networks in the European Union
• Bioinformatics 2025 Privacy-Preserving Federated Unsupervised Domain Adaptation with Application to Age Prediction from DNA Methylation Data
• arXiv preprint 2025 Accurate and Private Diagnosis of Rare Genetic Syndromes from Facial Images with Federated Deep Learning
• arXiv preprint 2025 Federated Learning for Epileptic Seizure Prediction Across Heterogeneous EEG Datasets
• IEEE TPS-ISA 2023 A Privacy-Preserving Framework for Collaborative Machine Learning with Kernel Methods

Explainable machine learning

Model explanations can expose sensitive training data. We study these privacy risks and design methods that provide insightful explanations while protecting the individuals whose data was used in training, making machine learning models more trustworthy.

Secure and trustworthy AI

We study how machine learning models can be attacked and defended: targeted adversarial attacks in black-box settings with few queries, defenses against model inversion attacks, and the geometry of image classifiers' decision regions. We also work on protocol-level safeguards for trustworthy agentic AI systems.

Key papers

• arXiv preprint 2026 Empirical Evidence for Simply Connected Decision Regions in Image Classifiers
• ICML 2026 Position: Trustworthy Model Context Protocol Enables Responsible Agentic AI!
• SaTML 2026 Accelerating Targeted Hard-Label Adversarial Attacks in Low-Query Black-Box Settings
• ICML Workshop 2024 Leveraging Multi-Color Spaces as a Defense Mechanism Against Model Inversion Attack