CV

Graduate Research Assistant

TruSe Lab, Purdue University

Leading research on membership inference attacks against fine-tuned LLMs and diffusion models under Professor Ninghui Li.

• Developed novel window-based MIA achieving 30% higher AUC than baselines
• Contributing to 5 papers at top security venues (ICLR, USENIX Security, NDSS)
• Co-developed SOFT framework reducing privacy leakage by 60% while maintaining 95% model utility

Undergraduate Research Assistant

Intelligent Systems Lab, RPI

Computer vision and robotics research under Professor Qiang Ji.

• Reduced MS-COCO object detection latency by 18% through novel backbone pruning
• Integrated real-time emotion/pose recognition on Pepper robot
• Successfully deployed system in 150-participant study

Undergraduate Research Assistant

Data Security and Privacy Lab, RPI

Pioneered new evaluation methodology for membership inference attacks under Professor Lei Yu.

• Published 2 papers (CCS'25, arXiv) on privacy attack research
• Co-developed MIAE toolkit with IBM Research featuring 8 attack algorithms
• Discovered critical vulnerability affecting 70% of existing MIA defenses

Undergraduate Research Assistant

Cognitive and Immersive Systems Lab, RPI

Research on LLM applications and creative AI under Professor Mei Si.

• Published 5 papers on creative AI applications (LREC-COLING, ICCC, AHFE, AIIDE)
• Developed prompt-chaining technique reducing GPT-3.5 hallucination by 40%
• Built multimodal pipeline deployed for 500+ users with 4.2/5 satisfaction rating

Ph.D.

Purdue University

Computer Science

• AI Privacy
• Large Language Models
• Membership Inference Attacks
• Differential Privacy

Bachelor of Science

Rensselaer Polytechnic Institute

Computer Science

• Machine Learning from Data
• Intelligent Virtual Agent
• Computational Creativity
• Machine Learning & Optimization

Dean's Honor List

Rensselaer Polytechnic Institute

Awarded for 6 consecutive semesters (Fall 2020 - Fall 2023)

Academic Recognition Letter

Charles V. Stewart, RPI

Outstanding academic performance recognition

Academic Recognition Letter

Mohammed J. Zaki, RPI

Outstanding academic performance recognition

Cascading and Proxy Membership Inference Attack

Network and Distributed System Security Symposium (NDSS)

Attack-agnostic framework incorporating membership dependencies via conditional shadow training to boost MIA performance.

Imitative Membership Inference Attack

USENIX Security Symposium

Novel imitative training technique requiring less than 5% computational cost of state-of-the-art approaches while outperforming existing MIAs.

Window-based Membership Inference Attacks Against Fine-tuned Large Language Models

USENIX Security Symposium

Introduced WBC method using sliding windows and sign-based aggregation, achieving 2-3x improvements in detection rates at low false positive thresholds.

Membership Inference Attacks on Finetuned Diffusion Language Models

International Conference on Learning Representations (ICLR)

First systematic investigation of MIA vulnerabilities in Diffusion Language Models. Introduced SAMA attack achieving 30% relative AUC improvement over baselines.

Evaluating the Dynamics of Membership Privacy in Deep Learning

arXiv preprint

Dynamic analytical framework for dissecting privacy leakage dynamics at individual sample level throughout training.

Membership Inference Attacks as Privacy Tools: Reliability, Disparity and Ensemble

ACM Conference on Computer and Communications Security (CCS)

Systematic investigation of disparities in MIAs through coverage and stability analysis, proposing ensemble framework for robust privacy evaluation.

SOFT: Selective Data Obfuscation for Protecting LLM Fine-tuning against Membership Inference Attacks

USENIX Security Symposium

Novel defense technique that mitigates privacy leakage by 60% while maintaining 95% model utility through influential data selection.

Reflections & Resonance: Two-Agent Partnership for Advancing LLM-based Story Annotation

Joint International Conference on Computational Linguistics (LREC-COLING)

Novel multi-agent system for automating story annotation, creating the 'StorySense' corpus with 615 annotated stories.

Enhancing Sentiment Analysis Results through Outlier Detection Optimization

arXiv preprint

Deep SVDD algorithm for outlier detection in text-based emotion analysis, improving performance across multiple datasets.

Prompt to GPT-3: Step-by-Step Thinking Instructions for Humor Generation

International Conference on Computational Creativity (ICCC)

3-stage template lifting humor ratings to 4.1/5 (+18%) while reducing token usage by 22%.

Visual Story Generation Based on Emotion and Keywords

AAAI Conference on AI and Interactive Digital Entertainment (AIIDE)

Story generation pipeline allowing user control over events and emotions, combining LLMs with diffusion models.

Automated Visual Story Synthesis with Character Trait Control

Applied Human Factors and Ergonomics (AHFE)

Novel technique for creating consistent human figures in visual stories using DreamBooth and Stable Diffusion.

MIAE: Membership Inference Attack Evaluation Framework

Comprehensive Python package with 8 state-of-the-art MIA algorithms and 3 evaluation metrics. Framework adopted by IBM Watson Research for internal privacy auditing of production models.

• 8 attack algorithms implemented
• Adopted by IBM Research
• Open-source contribution

LLM Science Exam - Kaggle Competition

Ranked top 19% (491/2,664) by fine-tuning GPT-J with LoRA and prompt ensemble techniques. Achieved 77% accuracy under 8GB GPU constraint.

• Top 19% ranking
• LoRA fine-tuning
• Resource-efficient implementation

Visual Story Generation Pipeline

End-to-end system combining LLMs with Stable Diffusion for interactive story creation. Deployed for 500+ users with 4.2/5 satisfaction rating.

• 500+ active users
• 4.2/5 user satisfaction
• Multimodal AI integration