Portfolio

Junbin Zhang (Thomas), Ph.D.

Welcome to my project showcase!
This repository serves as a central hub for exploring the 🎮games, 🔬research projects, 🛠️scripts, and 🌐websites I have developed.

🎮 Toy Arcade (Games)

A collection of fun games I have built.

Game	Code	Demo	Description	Highlighted Features	Tech Stack
DukeChess			♟️ Digital adaptation of the chess-inspired board game The Duke	- AI opponent - Cross-platform multiplayer support (desktop & web) - Docker image available for easy deployment	Desktop: Front-end: Back-end: Containerization:
MemoryDungeon			👾 Roguelike twist on the memory game	- AI opponent - Heroes with unique abilities - Gacha mechanics	Front-end (desktop): Back-end:
LinkWord			🔤 Lightweight Chinese word-chain game

ℹ️ If any of the demo links above do not work, please create an issue so I can take a look.

⤴️ back to top

🔬 Research Corner

A selection of my most meaningful research projects.

🔍 Image Forensics

Pixel-frequency Hybrid Approaches

• Shallow- & Deep-fake Images Localization
  
  
  A deep learning method that detects and localizes manipulated regions in both shallow- and deep-fake images.
  • Built upon the UperNet semantic segmentation framework.
  • A journal extension leveraging Vision Mamba and Graph Neural Networks is currently under review and available on arXiv.
  • An ONNX model converted from the original Python implementation is provided here, with custom functions implemented to support the conversion.

Frequency-domain Approaches

• Frequency-based AI-generated Images Detection
  
  
  A deep learning model that detects images generated by Stable Diffusion models using frequency-domain artifacts (>99% accuracy).

  • 🏰 Disney images were used as a case study.
  • A live demo is available here, with a backend running on AWS (API Gateway, Lambda, and S3) and the model exported to ONNX for efficient inference.
  • A Model Context Protocol (MCP) Server is implemented here as a Tool written in TypeScript, enabling LLMs to invoke the model for inference.

• Frequency-based AI-generated Images Attribution
  
  A frequency-based network that attributes fake images to their source Generative Adversarial Network (GAN) architecture (>95% accuracy).

Sensor-level Approaches

• AI-generated Images Detection Trained with Raw Images Only
  
  
  A Variational Autoencoder (VAE)-based approach trained solely on real/raw images for detecting AI-generated content (~94% accuracy).

• AI-generated Images Detection by Reconstructing Raw Images
  
  A method that detects AI-generated images by reconstructing Bayer patterns and comparing statistical differences between real and fake images (~98% accuracy).
  Notably, this method requires no deep learning.

📱 Mobile Security

• Comparison of Android Static Taint Analyzers
  
  
  A comparative study of Android static taint analysis tools (FlowDroid, Amandroid, and DroidSafe), evaluating their effectiveness and limitations using standardized benchmarks.

  • We built UBCBench, a benchmark suite for evaluating Android static taint analysis tools.
  • Bug fixes and improvements were merged by their original authors (details here).

• Android OAuth Vulnerability Analysis
  
  A course project that investigates security risks in Android applications using the OAuth protocol.

• Android Password Leakage Detection
  
  A course project that detects potential password leakage in Android applications using the FlowDroid framework, an Android static analysis tool.

🖼️ Image Generation

• TMO GAN
  
  
  A GAN designed to tone map high-quality 4K High Dynamic Range (HDR) images, significantly enhancing perceptual quality.

🥽 Virtual Reality

• Subjective Evaluation of Foveated Image Quality in Head-mounted Displays (HMDs)
  
  
  A novel subjective evaluation method for assessing the perceptual quality of foveated rendering in HMDs.
  • Developed an interactive voting mechanism using controllers in Unity to collect subjective user ratings.
  • Captured eye-tracking data via OpenVR and OpenGL and integrated it into the evaluation pipeline.

🎞️ Video Compression

• Visual-Attention-Model-based HDR Video Compression
  
  
  A visual-attention-driven HDR video compression method that incorporates perceptual saliency into the adaptive Quantization Parameter (QP) assignment of the High Efficiency Video Coding (HEVC) standard.

🎯 Serious Games

• BankPet
  
  
  A serious game–based banking application featuring a virtual character (🐱 everyone loves cats!) designed to help users develop better credit card and personal money management habits, validated through a user study.

⤴️ back to top

🛠️ Toy Workshop (Scripts)

Handy scripts I have crafted to make life easier.

🤖 MCP Servers

• Steam Review Summary
  
  An MCP server that analyzes and summarizes Steam game reviews, allowing LLMs to query review sentiment and generate concise summaries.

🐵 Tampermonkey Userscripts (JavaScript)

• Steam Wishlist Language Highlighter
  
  
  Highlights games in a user's Steam wishlist that lack support for specific languages.

  • Language information is retrieved using a Python script here as a GitHub Actions workflow.
  • A Google Chrome extension version is also available here.

• Arknights YiTuLiu Store Optimizer
  
  Annotates and 🏆prioritizes redeemed event resources on a popular third-party Arknights support site, helping players decide what to redeem first.

• ICBC Road Test Helper
  
  Simplifies the process of booking road tests on the Insurance Corporation of British Columbia (ICBC) website.

🐍 Python Scripts

• Steam Online Status Checker
  
  
  A tool using GitHub Actions to automatically monitor the status of a Steam user and send updates via Telegram messages.

• PRTS Scraper
  
  A lightweight script to scrape Arknights skill icons using MediaWiki APIs.
  These icons can later be used to train AI image generators to produce new icons.

💻 Code Contributions

• web-clipper
  
  
  I contributed to this Google Chrome extension that supports webpage clipping for popular note-taking apps (e.g., OneNote, Notion) by adding support for OneNote note groups.
  The pull request was merged by the original author.

⤴️ back to top

🌐 Website Creations

Websites I have built or contributed to.

• University of British Columbia, Digital Multimedia Lab
  
  
  I designed and currently maintain the website for my Ph.D. research lab, built on the Hugo Blox framework.
  I customized the framework to meet specific lab requirements:

  • Publications are categorized into finer-grained types (e.g., conference papers, journal articles), and the search function was rewritten to work within each category.
  • By adding new metadata to page elements, lab members can redirect to their own departmental web pages.
  • Embedded website statistics by modifying the universal footer.

• Personal Website
  
  
  My personal website built with HTML and jQuery.
  🌈 Refresh the page to see dynamically changing background colors!

• BC Liquor Store Explorer
  
  A course project web app that allows users to explore BC Liquor products with filters such as price, type, and country.

  • Highlighted features: interactive search and filtering, data visualization with Shiny.
  • Tech Stack: R, Shiny

⤴️ back to top

---

🧭 In the End: Why This Portfolio?

This repository serves two purposes:

• It presents an organized overview of my work, making my projects easier to understand and explore for readers beyond individual repositories.
• It also functions as a centralized index of my GitHub projects for long-term organization and reference.

As projects naturally grow across multiple repositories and organizations, GitHub currently offers limited support for presenting them as a coherent body of work. This portfolio serves as a developer-maintained overview that improves organization and discoverability, addressing a gap in how evolving personal GitHub project collections are presented.

⤴️ back to top