CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage
A coverage-curated dataset of 36,373 RGB-D frames across 1,275 indoor scenes, connecting panoramic perception with compact, geometry-consistent 3D environments.
Yudong Gao
I am a second-year Ph.D. student in Computer Science and Engineering at the Hong Kong University of Science and Technology (HKUST), advised by Prof. Shuai Wang.
I completed a combined B.Eng.–M.Eng. program at the China University of Petroleum (East China) under Prof. Honglong Chen, and was previously a Research Assistant at the College of William & Mary with Prof. Huajie Shao (2023–2024).
My research focuses on AI security, trustworthy language models and agents, and embodied intelligence.
Selected Research
Google Scholar ↗
PathMark: Protecting Intellectual Property of Mixture-of-Expert LLMs via Path Watermarks
Protects MoE LLMs through expert-routing watermarks, combining route alignment with contrastive learning. Verification reaches 99–100% across 4 models and 3 datasets, retaining 95% after 30 rounds of fine-tuning.
One Cognitive Loop Is Enough: SODA unlocks Pure-Text Spatial Reasoning in Large Language Models
Constructed 143,000 synthetic QA pairs and trained Qwen3-4B/14B with SFT and GRPO for pure-text spatial reasoning. Qwen3-4B accuracy improved from 34.5% to 76.8% on the SPOD-Bench CMM multi-turn task.
Bit-Flip Attacks on Vision-Language-Action Models: Action-Decoding Architecture Shapes the Vulnerability
Studies weight-integrity vulnerabilities across 4 VLA variants and 3 action-head families. Three selected logical INT8 bit flips reduced OpenVLA success from 88% to 0% in 50 LIBERO-Spatial simulation trials; also investigates localized protection.
SkillReducer: Optimizing LLM Agent Skills for Token Efficiency
Optimizes LLM Agent skills for token efficiency. Across 600 skills and 3,000 tasks, reduces description and body tokens by 48% and 39%, while improving the evaluated task score from 0.722 to 0.742.
A Triple Stealthy Backdoor: Hidden in Spatial, Frequency, and Feature Domains
Develops TriBA to hide backdoor triggers in the spatial, frequency, and feature domains, combining frequency-domain transformations with feature-space alignment. Evaluated on four datasets to study attack effectiveness and stealthiness.
Frequency Is Devious: Deep Stealthy Backdoor in Artificial Intelligence-Based Consumer Internet of Things
Develops DEST for AI-based consumer IoT, combining wavelet and Fourier transforms with singular-value embedding to construct strong triggers. Frequency restrictions conceal their traces, enabling stealthy backdoors in both spatial and frequency domains.
Energy-Based Backdoor Defense Without Task-Specific Samples and Model Retraining
Develops EBBA/EBBA+ to detect backdoored models without task-specific samples and remove backdoors without model retraining, using energy statistics and transferred-energy analysis.
Investigating the Backdoor on DNNs Based on Recolorization and Reconstruction: From a Multi-Channel Perspective
Studies backdoor attack and defense from a multi-channel perspective. RC-Attack embeds triggers in color channels, while RC-Defense reconstructs color and lightness channels to suppress poisoned behavior.
A Dual Stealthy Backdoor: From Both Spatial and Frequency Perspectives
Builds a dual-stealthy backdoor that hides triggers in both the spatial and frequency domains, studying how transformations affect trigger visibility and resistance to defenses.
Lightweight Privacy Protection Against Channel-Level Gradient Leakage Attack in Federated Learning
Studies channel-level gradient leakage in federated learning and develops lightweight privacy protection. Received the ICTC 2024 Best Paper Award.
Honors & Awards
- National Scholarship ×3
Ministry of Education · 2020, 2021, 2024 - Outstanding Master’s Thesis · 2025
- ICTC Best Paper Award · 2024
Academic Service
Journal ReviewerIEEE TIFS, IEEE TDSC, IEEE TNNLS, Pattern Recognition, Expert Systems with Applications, Engineering Applications of Artificial Intelligence
Conference ReviewerNeurIPS, ICLR, AAAI, WWW, ACM Multimedia
Sub-reviewerACM CCS, USENIX Security, FSE