Research in Action

Projects

Research projects spanning the lab's three thrusts, from active work to completed studies.

Current Research

Current Research Project

Current

Forensic Accountability in Heterogeneous Federated Learning

Research Problem

Federated learning can protect raw data locality while still remaining vulnerable to malicious client behavior. Existing Byzantine-robust aggregation methods primarily focus on suppressing or rejecting suspicious updates, but they do not inherently answer two forensic questions: which client was responsible for malicious behavior, and when did that behavior begin?

Why It Matters

In security-sensitive distributed learning systems, mitigation alone is not sufficient. Incident response, accountability, post-hoc investigation, and recovery require longitudinal evidence that can trace anomalous behavior back to a specific participant and localize the onset of an attack without confusing legitimate non-IID client variation with malicious activity.

Technical Focus

Longitudinal client-update forensics
Malicious-client attribution
Attack-onset localization
Non-IID-aware anomaly calibration
Magnitude and temporal-residual evidence
Persistent evidence accumulation
Robust aggregation versus forensic accountability

View project

Previous Research

Previous Research Projects

Privacy-Preserving Federated Medical Imaging with Encrypted Representations

Research Problem

Conventional federated learning avoids sharing raw medical images, but transmitted gradients or model updates can still leak sensitive patient information through reconstruction and inversion attacks. Encrypting entire gradients can also impose substantial communication and computation overhead.

Why It Matters

Cross-institutional medical AI requires strong privacy guarantees without making federated training impractical for bandwidth-constrained or privacy-regulated environments.

Technical Focus: Federated medical imaging, Vision Transformers, Homomorphic encryption, Gradient reconstruction resistance, Compact representation sharing, Secure aggregation, Communication-efficient privacy

View project

Privacy-Efficient Federated Learning for Distributed Medical Intelligence

Research Problem

Medical federated learning must balance privacy, model utility, and communication efficiency. Even when raw images remain local, model or gradient exchange creates privacy attack surfaces, while privacy-transformed or spatially altered inputs can degrade conventional CNN-based learning.

Why It Matters

A practical privacy-preserving medical AI system must remain useful under constrained bandwidth and privacy-preserving transformations while minimizing exposure of sensitive training information.

Technical Focus: Privacy-first federated learning, Compact ViT embeddings, Differential privacy, Medical image analysis, Communication reduction, Privacy-transformed inputs, Distributed learning efficiency

View project

Human-Centered Evaluation of LLM-Generated Code Quality

Research Problem

Human evaluation of LLM-generated code is often inconsistent across subjective quality dimensions such as readability, modularity, style, understandability, and effort. Ambiguous evaluation guidelines can reduce reproducibility and make benchmark results difficult to compare.

Why It Matters

Reliable evaluation is essential for understanding whether AI-generated code is genuinely useful, maintainable, and aligned with human expectations rather than merely functionally correct.

Technical Focus: LLM-generated code evaluation, Human-in-the-loop AI, Rubric-guided assessment, Inter-rater reliability, Multi-language code quality, AI-assisted feedback, Reproducible software-quality evaluation

View project