Research depth that supports practical AI/ML work.

My research asks a practical question: when machine learning is placed inside a real system, does it still behave reliably when the data are noisy, incomplete, adversarial, or physically constrained? I study this question primarily through hardware security, where ML models must reason over side-channel measurements, process variation, limited labels, and attackers who may deliberately adapt to the detector.

A major part of my work focuses on ML-assisted hardware Trojan detection and localization. Semiconductor systems are increasingly exposed to untrusted design, fabrication, testing, and integration environments, so detection methods cannot depend only on clean benchmark accuracy or ideal golden references. I build and evaluate pipelines that use side-channel evidence such as Ring Oscillator Network, power, frequency, and electromagnetic behavior to identify malicious or anomalous circuit behavior.

My recent work pushes this direction toward adversarial robustness. In the NSF REU project I mentored and helped develop, we showed that a detector with strong nominal performance can fail under gradient-based adversarial perturbations. We then studied synthetic data augmentation strategies, including SMOTE, CTGAN, and TVAE, to make hardware Trojan detection models more resilient. This line of work reflects the larger principle behind my research: secure AI systems need evaluation under stress, not only evaluation under friendly conditions.

Beyond hardware security, I apply the same research lens to LLM-assisted explainability, rubric-aware assessment, hyperdimensional computing for IoT anomaly detection, and scientific ML for drug-target affinity prediction. The domains differ, but the core goal remains consistent: build AI/ML systems that are reproducible, interpretable enough for human use, efficient enough for deployment, and robust enough to be trusted when the environment changes.