I am Haoze Yu (于浩泽), an undergraduate student majoring in Computer Science & Technology at Harbin Institute of Technology, Shenzhen (HITsz), expected to graduate in June 2027.

My current research focuses on multimodal foundation models, large language model security, adversarial robustness, privacy-preserving RAG, and AIGC/deepfake detection. I am interested in understanding the failure modes of multimodal AI systems and developing practical methods for attack analysis, defense, traceability, and trustworthy deployment.

• 🎓 B.Eng. in Computer Science & Technology, HITsz, 2023–2027 (expected)
• 📈 Weighted Average: 93.88 / 100
• 🏅 Grade Rank: Top 4.5%
• 📚 Major Rank: Top 5.4%
• 🔬 Research interests: Multimodal & LLM Security · VLM Red-Teaming · Adversarial Attacks · RAG Privacy · AIGC Deepfake Detection

I study the robustness and security boundaries of vision-language models and multimodal ranking/reasoning systems, with a particular focus on adversarial attacks against visual inputs. My work includes optimization-based red-teaming, robustness evaluation, attack transferability, and efficient multimodal inference during iterative adversarial testing.

I am exploring the privacy–utility trade-off in RAG systems when retrieved corpora contain sensitive or personally identifiable information. My work involves benchmark construction, privacy/utility evaluation, sensitivity analysis, and training-free safety mechanisms that can be applied to both open-source and closed-source language models.

I work on audio-visual forgery detection and traceability, aiming to move beyond binary real/fake classification toward identifying where the forgery occurs and why the model considers it suspicious. My research combines appearance, motion, style, audio-visual synchronization, and reliability-aware evidence fusion.

I am also interested in LLM-based agents, especially evaluation harnesses, capability evolution, memory/skill systems, and methods for making iterative self-improvement more stable and measurable.

Project Lead · Harbin Institute of Technology, Shenzhen
Oct 2025 – Mar 2026

• Designed a multi-stage framework for multimodal forgery analysis using visual appearance, optical-flow-based motion cues, style evolution, and audio-visual consistency.
• Focused on jointly producing real/fake prediction, forged-region localization, and interpretable cause analysis.
• Investigated reliability-aware fusion across heterogeneous forensic signals.

Project Lead · Harbin Institute of Technology, Shenzhen
Dec 2024 – Jul 2025

• Designed a hierarchical framework for encrypted and malicious traffic analysis.
• Combined lightweight screening, coarse attack-family classification, fine-grained attribution, routing decisions, and probability-calibrated fusion.
• Focused on balancing detection efficiency, classification accuracy, and attribution granularity.

• Three National-level Scholarships
  • National Scholarship (2025)
  • National Encouragement Scholarship (2024, 2026)
• First-class Academic Scholarship, Harbin Institute of Technology (2025)
• Top 10 Youth Volunteers, HIT (2025) — only 10 selected university-wide per year
• National Second Prize, National AI Competition for College Students
• Provincial First Prize, China Undergraduate Mathematical Contest in Modeling (CUMCM)
• First Prize, Chinese Mathematics Competitions (CMC)
• Excellent Social Practice Member (2025)
• Excellent Student (2023–2025)
• Excellent Youth League Member (2023–2025)
• 1 registered software copyright

Programming
Python · C++ · C · Java

Machine Learning / Deep Learning
PyTorch · Transformer architectures · Model training · Hyperparameter tuning · Evaluation & analysis

LLM / Multimodal Practice
Fine-tuning · Inference · Deployment · RAG pipelines · Multimodal models · Safety evaluation

Security Research
VLM red-teaming · Adversarial attacks · Robustness analysis · PII privacy protection · AIGC deepfake detection & traceability

Systems & Engineering
Linux · Git · Experiment reproduction · Scripting · Collaborative development · RISC-V CPU design

Researching secure, reliable, and interpretable multimodal AI systems.