Backdoor Attacks in Large Language Model Reasoning
Investigating a novel backdoor attack targeting the reasoning process of large language models.
Kennesaw State University, USA · Supervisor: Dr. Honghui Xu · July 2026 – Present
Overview
Investigating how the reasoning process of large language models can be compromised. The work combines chain-of-thought analysis with trigger-injection techniques to identify where reasoning chains are vulnerable and how corrupted intermediate steps propagate to final answers.
Approach
1
Attack design
Proposed a novel backdoor attack targeting the reasoning process of large language models.
2
Training pipeline
Designed the training process used to produce the corrupted model.
3
Evaluation
Performed intensive evaluation against prior work, and tested the robustness of the proposed attack against state-of-the-art defenses.
4
Writing & organization
Helped organize the paper’s structure and contributed to writing.
Status
One paper is in preparation.
Tools
Python PyTorch Hugging Face Transformers