Research Focus

Robustness & OOD

AI systems resilient to unexpected data, adversaries and environmental changes

What is Robustness in AI?

A robust system continues to function reliably even when facing data or situations it has not seen during training. Out-of-Distribution (OOD) refers precisely to data outside the expected range.

We investigate defenses against adversarial attacks, methods to detect when a model faces OOD data, and continuous adaptation techniques that allow the system to update in dynamic environments.

Sources of Vulnerability

Different types of robustness challenges

Adversarial

Adversarial Attacks

Small perturbations in input can cause incorrect predictions. We research defense against pertinent attacks and certified robustness.

Distribucional

Distribution Shift

Data distribution in production differs from training. Adaptation methods without access to true labels.

Ambiental

Environmental Changes

The operating environment changes: sensors age, conditions vary. Systems that adapt continuously without losing prior knowledge.

Defense Strategies

Techniques to improve robustness

Adversarial Training

Train on adversarial examples so the model learns to be resilient to perturbations.

OOD Detection

Identify when data is out-of-distribution to refrain from making unreliable predictions.

Continuous Adaptation

Update the model dynamically with human feedback or new data without catastrophic forgetting.

Current Challenges

Open problems in robustness

Certified Robustness

It is difficult to formally guarantee that a model is robust. We seek verification methods that provide robustness certificates.

Robustness-Utility Trade-off

Improving robustness often reduces accuracy on normal data. We research methods that balance both objectives.

Scalability

Robustness methods are computationally expensive. We need techniques that scale to large models.

Practical Applications

Domains where robustness and reliability are non-negotiable

Autonomous Vehicles: Ensuring perception and control systems remain safe under adverse weather, sensor failure, and unexpected road conditions.
Healthcare Diagnostics: Building medical imaging and signal-analysis models that maintain accuracy when data distribution shifts across hospitals or equipment.
Financial Systems: Protecting fraud-detection and trading algorithms against adversarial manipulation and market anomalies.
Critical Infrastructure: Securing power grids, water networks, and transportation systems against AI-driven cyberattacks and sensor noise.
Robotics: Enabling robots to operate reliably in unstructured environments where training conditions differ from deployment scenarios.
Biometric Security: Defending face-recognition and fingerprint systems against spoofing attacks and out-of-distribution presentation attempts.

Our Research

In the Robustness and OOD line of the Trustworthy AI group at IAFER, we are investigating systems that maintain reliable performance beyond their training distribution, critical for safety-critical applications.

🚀 Active Research Lines

Explainable Out-of-Distribution Detection: Statistical methods that identify anomalous samples in a transparent and certifiable manner
Robustness to perturbations: Techniques to train models resistant to adversarial attacks and domain changes
Robustness evaluation in biomedical images: Validation of computer vision systems in medical contexts with quality guarantees
Generation of robust representations: Learning methods that produce features invariant to irrelevant variations

📚 Publicaciones

Loading publications...