Research Focus

Privacy & Federated Learning

Distributed learning with differential privacy guarantees and trust mechanisms

What is Federated Learning with Privacy?

Federated Learning (FL) allows training AI models without centralizing sensitive data. Data remains on local devices while only model updates are shared.

We complement FL with differential privacy to ensure that even those updates do not reveal sensitive information about individuals. Our work on distributed trust mechanisms ensures that participants can trust the process without centralized intermediaries.

Key Components

Fundamental elements of a trustworthy and private FL system

Distribuido

Federated Learning

Train models without centralizing data. Local clients train on their private data and send only gradients to the coordinating server.

Privacidad

Differential Privacy

Add controlled noise to gradients to mathematically guarantee that individual data cannot be reconstructed.

Confianza

Trust Mechanisms

Blockchain, homomorphic encryption and cryptographic verification to ensure that no participant can falsify results.

Use Cases

Domains where private FL is essential

Mobile Devices

Train models on phones without sending personal data to central servers.

Distributed Medical Data

Train clinical models with data from multiple hospitals while keeping medical secrets.

Financial Institutions

Collaboration between banks without revealing sensitive customer data or competitive information.

Current Challenges

Open problems in private and trustworthy FL

Convergence Under Privacy

Differential privacy adds noise that slows convergence. We seek methods that maintain privacy guarantees without sacrificing too much performance.

Data Heterogeneity

Data distributed among participants is non-IID (not identically distributed). Training is slower and more unstable.

Attacks in FL

Gradients can reveal information even with differential privacy. We develop robust defenses against reconstruction and inversion attacks.

Practical Applications

Scenarios where privacy-preserving AI is essential

Healthcare Collaboration: Hospitals jointly train diagnostic models on distributed patient data without sharing sensitive records, enabling multi-center studies while preserving patient confidentiality.
Financial Fraud Detection: Banks collaborate to detect cross-institutional fraud patterns without exposing individual transaction histories or customer identities.
Smart Cities: Federated traffic and energy optimization across municipalities where local sensor data remains private but collective models improve urban planning.
Mobile Personalization: On-device keyboards and assistants learn user preferences from local interactions without uploading raw keystrokes or voice recordings to central servers.
Pharmaceutical Research: Competitors and research institutes train drug-discovery models together on proprietary compound data while keeping molecular structures and trial results strictly confidential.
Wearable Health Monitoring: Fitness and medical wearables train models on user biometrics locally, contributing only encrypted parameter updates to improve global health models.

Our Research

In the Privacy + Trust in Federated Learning line of the Trustworthy AI group at IAFER, we are developing secure methods for machine learning collaboration without centralizing sensitive data.

🚀 Active Research Lines

Federated learning with differential privacy: Train distributed models with mathematical privacy guarantees
Distributed trust mechanisms: Systems that verify participant integrity without central trust
Defense against membership attacks: Protection against inference of whether a data point was used for training
Secure multi-institutional collaboration: Frameworks for institutions (hospitals, banks) that train jointly

📚 Publicaciones

Loading publications...