Machine learning researcher · Virginia Tech
I develop machine learning methods for recognizing anomalies that differ from those seen during training. My research connects anomaly detection, out-of-distribution generalization, and adversarial robustness, with applications in cybersecurity, multivariate time series, and cyber-physical systems.
At Virginia Tech, I study how useful structure in data—relationships between variables, shared information across domains, and the geometry of learned representations—can help models recognize unfamiliar behavior. I am also interested in large language models and in-context learning for zero-day anomaly detection.
ANOMALY Detection · Generalization · Robustness
A detector must transfer what it learns from familiar data to recognize new anomaly categories. This framework first learns normal behavior, then uses a few examples of known anomalies to calibrate confidence. Repeating the process across training tasks helps distinguish normal observations from previously unseen attacks or abnormal heartbeats.
A system can behave abnormally even when each sensor looks plausible on its own. This work combines a transformer for temporal context with explicit models of how latent variables depend on one another. Contrastive training separates normal windows from anomalous ones, using a small labeled anomaly pool to create realistic perturbations.
Training on known attacks can teach a detector patterns that fail to transfer. This work combines classification and reconstruction with an information penalty on a compact latent representation. The aim is to retain useful detection evidence while reducing domain-specific correlations, improving generalization to anomaly categories held out during training.
Poisoned training data can undermine regression models used in smart-grid monitoring. This paper uses discrepancies between predictions and observations to separate candidate clean and corrupted subsets through a two-level optimization. It examines varied attack strategies and contamination levels, including majority-poisoned settings, while discussing the effects of temporal autocorrelation.
Distributed-energy controllers can be attacked both by corrupting their training data and by disguising malicious commands at deployment. This work couples robust regression with detection-threshold selection in a bilevel learning problem, using local measurements to check expected commands. Experiments examine combined attacks on an IEEE 13-bus feeder.
The first three authors contributed equally.
An autoencoder for skewed sensor data combines reconstruction error with a robust Mahalanobis measure in latent space. Using median- and MAD-based statistics helps account for correlations that Euclidean reconstruction error misses, with the goal of separating anomalies close to normal observations as well as more distant anomalies.
Machine-learning systems introduce data-security risks across collection, storage, transmission, and computation. This survey connects established cybersecurity practices with privacy-preserving machine learning, reviewing access controls, cryptography, zero trust, differential privacy, and federated learning. It explains assumptions and trade-offs to help practitioners select protections for a system’s data lifecycle.
ARM3Dnet forecasts county-level COVID-19 outcomes by combining mobility-informed graph diffusion with autoregressive sequence modeling. A mixture-density output represents multiple plausible outcomes rather than a single deterministic forecast. The study evaluates how changing mobility relationships and probabilistic modeling affect forecasting accuracy and uncertainty on county-level epidemic data.
Training & experience
Virginia Tech
Virginia Tech
Focus in machine learning
Virginia Tech
Machine Learning and Advanced Machine Learning
Accenture AI
Statistical learning, production recommendation systems, and interpretable models
Accenture Digital
Data analytics, cloud microservices, and deployment automation
Open to conversations about machine learning opportunities.