Machine learning models are now used in many important areas, from healthcare and finance to autonomous systems and natural language processing. However, the robustness of these models remains a critical concern: how will they behave when confronted with adversarial examples, distribution shifts, corrupted or missing data, or subtle implementation errors? Without strong robustness guarantees, ML systems risk failure, liability, and lack of trust.

Our lab focuses on the robustness analysis of machine learning models, developing methods to measure, guarantee, and improve the performance of ML systems under real-world perturbations and uncertainties. The research is at the intersection of theoretical foundations (statistical learning theory, verification, optimization) and applied systems (model deployment, adversarial attacks, safety-critical applications).

• Adversarial and Perturbation Analysis

Using techniques from adversarial ML, we examine how models respond to worst-case perturbations (e.g., small input changes that lead to misclassification), and develop methods to certify robustness bounds.

• Verification & Formal Guarantees for ML Models

We explore algorithmic frameworks that provide formal guarantees about model behavior (e.g., bounded error under bounded perturbations), drawing on verification, formal methods, and program analysis.

• Robustness in Deployment 

Beyond standalone models, we study how ML components interact with other system parts and how robustness must be maintained end-to-end in real applications.