Projects
Research Projects
EviGuard: Cross-Modal Safety Unlearning in Multimodal LLMs
Jan 2026 – Apr 2026
Evidence-guided connector intervention for cross-modal safety unlearning in multimodal large language models.
- Identified that cross-modal risk in multimodal LLMs concentrates in a low-rank subspace at the visual-language connector, motivating targeted intervention rather than uniform suppression
- Reduced over-refusal rate (SARR) from 30.3% to 22.3%; on OOD benchmark SIUO, achieved 9.8% ASR and 68.0% Safe & Effective rate vs. best baseline (81.2% / 8.0%)
- Contributions: Experimental framework design, algorithm implementation, ablation analysis, manuscript writing
BEACON: Budget-Efficient Discovery of Policy Violations in LLMs
Nov 2025 – Jan 2026
Budget-efficient discovery of policy violations via cognitive-guided Monte Carlo Tree Search.
- Reframed LLM safety evaluation as budget-constrained failure discovery; formalized efficiency-oriented metrics (k-FDQ, NDA, CCR, DV) capturing discovery timing and harm category diversity beyond traditional ASR
- Built Cognitive-Guided MCTS with defense persona profiling and diversity-aware selection; achieved 85.5–100% ASR across 6 frontier LLMs, discovering failures 3.7× faster than strongest baseline (k-FDQ 26 vs. 95)
- Contributions: Problem formulation, attack design, experimental analysis, manuscript preparation
VidTouch: Vision-based Tactile Perception via Multimodal Learning
Dec 2024 – June 2025
A multimodal benchmark dataset for dynamic visuo-tactile fabric recognition.
- Built VidTouch (145 fabric categories, 440 RGB images / 440 tactile videos), the first dynamic multimodal benchmark for fabric recognition; supports multi-label classification, cross-modal retrieval, and zero-shot generalization
- Designed X3D + EfficientNet + MLP fusion pipeline achieving 98.6% in-category accuracy; revealed zero-shot generalization bottleneck (41.3%), positioning VidTouch as an open challenge benchmark
- Contributions: Dataset construction, model training and optimization, and experimental validation
AttentiveCSI: WiFi-based Gesture Recognition with CNN-LSTM
June 2025 – Aug 2025
WiFi-based Channel State Information gesture recognition using attention-enhanced CNN-LSTM network at NC State University.
- Developed AttentiveCSI system for non-invasive gesture recognition using WiFi CSI data, eliminating the need for wearable sensors or cameras
- Implemented attention-enhanced CNN-LSTM architecture achieving high accuracy in recognizing complex hand gestures across different environments
- Awarded First Place in Poster Symposium at U.S. GEARS Research Program (1/50 participants)
- Supervisor: Dr. Muhammad Shahzad, NC State University
Autonomous Driving Decision-Making based on Game Theory
Mar 2024 – Feb 2025
Lane change decision-making method for intelligent vehicles in foggy weather using dynamic game theory.
- Designed a decision-making framework for intelligent vehicles operating in foggy weather conditions, utilizing dynamic game theory to model interactions between vehicles
- Implemented and validated the system using CarSim & Simulink simulations, demonstrating safe and efficient lane-change decisions under reduced visibility scenarios
- Result: Chinese Invention Patent ID: 202510141836.1 (Passed Preliminary Examination)
- Advisor: Dr. Heng Deng, Beijing University of Technology