About

I am an MPhil researcher in Computer Science and Engineering at the University of New South Wales, advised by Dr. Jiaojiao Jiang. My research focuses on the integrity and security of generative models — how AI-generated content can be watermarked, how those watermarks can be attacked, and how manipulated media can be detected.

Generative Model Watermarking Watermark Removal Attacks Deepfake Detection Diffusion Models Robust Graph Learning Large Language Models

News

  • Aug 2026 The official CSI-Attack code for our WWW 2026 paper is now available.
  • Aug 2026 SLICE was accepted as an Oral Presentation at BMVC 2026. 🎉
  • Aug 2026 New preprint IRIS, on visual-semantic binding for forgery-resistant watermarking of diffusion images.
  • Jul 2026 The TRACE project page is live, with an interactive demo of tamper-evident attribution for AI-agent trajectories in health, safety and environment workflows.
  • Jul 2026 BIP was accepted at ECCV 2026, and LAVA was accepted at ACM Multimedia 2026. 🎉
  • Jul 2026 Two new preprints: TRACE, on attribution watermarks for LLM-agent trajectories, and Watermark Forensics, an information-theoretic view of watermark forensics.
  • 2026 GSCNet is out in the Journal of Visual Communication and Image Representation, vol. 120, art. 104895.
  • Jun 2026 New preprint Flood and Harvest, on the necessity of trivia in language generation for mathematics.
  • Apr 2026 Our paper Breaking Semantic-Aware Watermarks was presented at The Web Conference (WWW) 2026.
  • Mar 2026 Two new preprints on diffusion-model watermarking: SHIFT and SLICE.
  • 2025 Started my MPhil at UNSW Sydney.

Publications

2026

2025

2024

Projects

CSI-Attack

LLM-guided robustness testing for semantic-aware image watermarks

Official research code accompanying our WWW 2026 paper. The repository provides calibrated evaluation for three semantic-aware watermark schemes, a plug-in interface for additional methods, and a CPU-only self-test.

Code Paper arXiv

TRACE

Tamper-evident attribution for AI-agent trajectories

Two complementary watermark channels stamped onto an agent's decision trajectory, so the record still proves which agent produced it even after steps are deleted or rewritten. The project page walks through two health-and-safety scenarios and includes a live demo.

Live demo Code Paper

Education & Experience

  • 2025 – present

    MPhil in Computer Science and Engineering

    University of New South Wales, Sydney, Australia

    Advised by Dr. Jiaojiao Jiang. Research on generative model watermarking, watermark removal attacks, and deepfake detection.

  • 2023 – 2024

    Undergraduate Researcher

    City University of Macau, Macau, China

    Developed G-RXAD, a deep reinforcement learning-based attack detection model for network security, published at AICIT 2024.

  • 2021 – 2025

    BSc in Data Science

    City University of Macau, Macau, China

    Foundation in machine learning, data-driven systems, and applied AI research.