Our machine learning work runs from theory to production: what can be learned, how you verify that it was, and what it costs to run.
Theory
The research side works in statistical learning theory: PAC learning and, in particular, verification. What does it take to check that a model learned what it claims, with less work than training it yourself? That question matters to anyone who buys, audits, or regulates a model they didn’t build.
Practice
The applied side ships models whose decisions can be explained and audited: bias testing under the NIST AI Risk Management Framework, drift and anomaly detection in production, and retrieval-grounded systems that cite their sources instead of improvising.
Adversarial ML
We also work on media authentication (telling generated content from real), including deepfake-detection research by our team published at ACM CCS 2025.