Soohyeon Choi

Soohyeon Choi

AI and LLM Safety and Security Researcher

I am Soohyeon Choi, a Research Scientist at Singapore Management University. I received my Ph.D. in Computer Science from the University of Central Florida in 2025, advised by Prof. David Mohaisen. My research asks how advanced language models behave under adversarial and deployment shifts and how AI-generated code can be detected, attributed, and provenance-marked.

I work across compositional jailbreaking, exposure-aware generated-code detection, human and AI-agent code attribution, and post-hoc code watermarking. I build reproducible benchmarks and evaluation pipelines across Python, Java, and C++ using fine-tuned code encoders, zero- and few-shot LLMs, matched controls, robustness testing, and statistical auditing. I previously served as a Research Officer at A*STAR’s Institute for Infocomm Research in Singapore.

Research Interests

  • AI and LLM safety and security
  • Adversarial evaluation and deployment shifts
  • Jailbreaking and model-behavior analysis
  • Human and AI-agent code attribution and provenance
  • Code watermarking and generated-code detection

Current Research

Talking to Themselves: An Exposure-Aware Evaluation of LLM-Generated Code Detection. We evaluate heuristic and LLM-based generated-code detectors under data exposure, low-effort adversarial edits, and model-level attribution across C++, Java, and Python. This manuscript is under review.

I Can Find You in Seconds! Leveraging Large Language Models for Code Authorship Attribution. We develop zero-shot, few-shot, and tournament-style LLM methods for source-code authorship attribution across C++ and Java, including large candidate sets and adversarial transformations. This manuscript is under review. Preprint

Compositional Jailbreaking of Aligned LLMs. We study how simple jailbreak transformations interact when applied sequentially, including synergistic, destructive, and structurally incompatible compositions. This manuscript is under review. Preprint

CommitTrace. We study code-only attribution of AI coding agents from repository-local file patches and hunks. The work evaluates prior-method transfer, model robustness, cross-language generalization, and repository-disjoint deployment shifts. This manuscript is under review.

Multi-Channel Spread-Spectrum Code Watermarking. We develop a post-hoc, training-free method that embeds a 24-bit provider label through variable-naming and structural channels with explicit recovery assumptions. This manuscript is under review. Preprint

Education

Ph.D., Computer Science, University of Central Florida, Orlando, FL, USA (2021 – 2025)

  • Advisor: Prof. David Mohaisen
  • Topic: Machine Learning, AI, LLM & Security

M.Sc., Computer Science, South Dakota State University, Brookings, SD, USA (2018 – 2021)

  • Advisor: Prof. Sung Shin
  • Topic: Sensor Network, Authentication & Security

B.Sc., Computer Engineering, Keimyung University, Daegu, South Korea (2011 – 2017)

Professional Experience

  • 05/2025 – Present: Research Scientist, Singapore Management University – AI, LLM, and machine-learning security

  • 08/2021 – 05/2025: Graduate Research Assistant, University of Central Florida – machine learning, LLMs, and security

  • 12/2023 – 12/2024: Research Officer, A*STAR Institute for Infocomm Research – machine learning and security

  • 08/2019 – 12/2020: Graduate Research Assistant, South Dakota State University – sensor-network security

  • 08/2018 – 05/2019: Teaching Assistant, South Dakota State University – CSc 300 Data Structures