LATEST WORK
MechAudit-40: White-Box Auditing across 40 LLM Attack Mechanisms
A systematic study of internal attack signatures across 40 mechanisms, with a runtime auditor evaluated on previously unseen attacks.
SAINT LOUIS UNIVERSITY
Zhen Guo.
I study AI security and trustworthy AI, with a focus on reasoning backdoors, model auditing, and privacy. I am supervised by Dr. Reza Tourani.
PAPERS & PREPRINTS
Research timeline.
Research in AI security, privacy, and communication systems.
2026
2 papers-
MechAudit-40 · Discovery-to-auditing pipelineView full size TraceGuard: Process-Guided Firewall against Reasoning Backdoors in Large Language Models
A compact reasoning firewall that audits intermediate steps and identifies where a reasoning trace first becomes unsupported.
TraceGuard · Three-stage training pipelineView full size
2025
2 papersPersistent Backdoor Attacks in Continual Learning
Investigating how backdoors survive continual model updates through Blind Task Backdoor and Latent Task Backdoor attacks.
Persistent Backdoors · Attack overviewView full size DarkMind: Latent Chain-of-Thought Backdoor in Customized LLMs
Exposing latent backdoors that activate within a model’s chain of thought to alter its output without changing the user’s query.
DarkMind · Latent reasoning backdoor pipelineView full size
2026
Communication systemsUnleashing the Power of Vibration and Acoustic Signals for Fast and Reliable Communication
Vib-Sound combines vibration and acoustic signals with collaborative cross-channel demodulation for communication between commodity devices.
Vib-Sound · Cross-channel demodulation architectureView full size
CONTACT
For questions about my work or research collaborations, please get in touch.