Experience

Work Experience

• Syntiant Corp. · Engineering Intern · Irvine, California (2022.7 – 2022.9)
  • Developed a confidence-aware, multi-teacher knowledge distillation framework for keyword spotting tasks, leveraging the student-teacher architecture.
  • Used pre-trained transformer models to enhance student model performance effectively.
• Ericsson Inc. · Data Science Intern · Santa Clara, California (2020.6 – 2021.1)
  • Built a secure federated XGBoost framework with an innovative secure quantile sketch and practical secure aggregation.
  • Implemented pairwise secret masking of model parameters to protect against data leakage during aggregation, strengthening client data privacy.

Teaching Experience

• UC Irvine EECS Department · Teaching Assistant (Winter and Spring 2024; Winter and Spring 2025)
  • Taught EECS 31L lab sessions on Verilog design, simulation, testbench creation and waveform debugging.
  • Held office hours, online Q&A, provided one on one debugging help, code reviews, and study guidance.
  • Refined lab materials and grading rubrics with the teaching team; evaluated reports and Proctored exams.

Research Experience

1. Agentic AI and LLM

• ReTalk Agent: Multilingual Video Dubbing, 2025.8 – 2025.10.
  • Developed an AI agent pipeline for multilingual video dubbing using LLM-based language processing while preserving speaker voice identity.
  • Implemented lip-synchronization and voice cloning for natural facial-speech alignment.
• LLM-Powered Voice Assistant with Voice Cloning, 2025.6 – 2025.8.
  • Designed an interactive voice assistant using LLMs as the core reasoning engine for natural, context-aware conversations.
  • Integrated voice cloning for personalized synthetic voice, enabling seamless spoken interaction and enhanced user engagement in real-time dialogue.

2. Client Valuation and Selection

• Valuing Solo and Synergy in Federated Learning, 2024.10 – 2025.7.
  • Proposed DuoShapley, an efficient Shapley value approximation that adaptively balances individual and collaborative contributions in heterogeneous FL settings.
  • Demonstrated low overhead and higher robustness with noisy users, enabling scalable client selection for large-scale FL.
• Rewarding the Rare: Maverick-Aware Shapley Valuation in Federated Learning, 2023.10 – 2024.9.
  • Designed a Maverick-aware Shapley valuation framework to quantify client contributions under data heterogeneity, addressing undervaluation of clients with rare classes.
  • Developed FedMS, a Shapley-guided client selection mechanism that adaptively prioritizes high-contribution clients, improving global performance and robustness against adversaries and free-riders..

3. Model Pruning (Efficiency & Privacy)

• PriPrune: Quantifying and Preserving Privacy in Pruned Federated Learning, 2022.11 - 2023.8.
  • Derived information-theoretic upper bounds on information leakage in pruned FL models.
  • Developed PriPrune, a privacy-aware pruning algorithm with personalized per-client defense masks and adaptive pruning rates, balancing privacy and performance.

4. Privacy-Preserving Federated Machine Learning

• Information Leakage in Personalized Federated Learning, 2022.4 – 2022.6.
  • Executed gradient leakage attacks (DLG) in personalized FL, demonstrating how personalization levels impact vulnerability.
  • Proposed PerFed-LDP, a per-example differential privacy method for personalized FL, analyzing privacy-utility tradeoffs.
• Privacy by Projection: Federated Population Density Estimation, 2021.12 – 2022.7.
  • Designed a federated kernel density estimation (KDE) framework to estimate population density while keeping data local.
  • Developed a federated Random Fourier Feature (RFF) KDE approach projecting user data onto spatially delocalized basis functions, enhancing privacy without compromising accuracy.
• Location leakage in federated signal maps, 2021.11 – 2022.4.
  • Executed gradient leakage attacks on federated signal map prediction, reconstructing average locations from private spatio-temporal datasets.
  • Proposed batch selection defenses to obfuscate true locations, balancing mapping utility with privacy protection.
• SaferQ: Obfuscating Search Queries via Generative Adversarial Privacy, 2020.3 – 2020.5.
  • Developed SaferQ, extending the GAP framework for sequence generation to obfuscate search queries while balancing privacy and utility.
• Secure Federated XGBoost Framework, 2018.9 – 2019.8.
  • Designed a secure federated XGBoost framework with anonymized data aggregation to balance privacy and performance.