I work on machine learning for security at UTRGV, with a focus on federated learning and network detection.
Benign-behavior transfer in federated network detection
Organizations want to train network anomaly detectors together without handing over raw traffic. Federated learning allows that, but sharing a model also shares each site’s idea of what normal traffic looks like. My current work measures when that shared benign behavior helps a detector and when it quietly hurts one, the failure mode known as negative transfer.
The study uses a three-organization testbed with verified traffic labels and compares federated training against local-only and shared-encoder baselines, scoring each on detection recall against false alerts per host-day.
- Live progress: flndr.damianvillarreal.com
- Advisor: Dr. Sergei Chuprov
- Target: conference submission, December 2026
Does filling sensor gaps help intrusion detection?
Network sensors sometimes stop reporting, which leaves gaps in the data an intrusion detector depends on. I tested whether an AI model that fills those gaps (CSDI, a diffusion model) helps catch attacks better than simpler fixes.
So far the answer is no. A detector trained to expect gaps does as well or better at a fraction of the cost, and the gap filler ran about 200 times slower at inference while still losing on the sealed test. The useful takeaway is a negative one: teach the detector about missing data instead of repairing the data first.
- Course research project, CSCI 6379
- Tested on sealed CTU-13 scenarios under rules fixed before scoring
Federated pre-training for security-assistant LLMs (FeDAPT)
Can a group of security organizations train a better security-assistant LLM together without any of them sharing private telemetry? FeDAPT splits the problem. The shared part is public security prose (ATT&CK, CVEs, Sigma rules), which organizations federate into a shared adapter. Each org’s raw logs stay local and are used only for a second, local tuning stage.
The payoff of joining the federation is a single number: the gap between federated adaptation and each org going it alone. The system is built and tested, the data layer is validated on real public logs, and the experiment is staged for an HPC cluster. The full training run is the next step, so there is no headline result yet.
- Summer 2026 graduate research assistant work
- Code: github.com/dsuyu1/fedapt