Publications

ICICS 2026Accepted · Conference paper

Small Models, Big Variances: Unveiling the Impact of Execution Environments on SLMs’ Security

Zijun Li, Tengfei Tu, Yuxin Lian, Zicheng Kong, Yiling Zhou, Aofan Liu, and Wenmin Li

SecuDevSLM evaluates small language model security under mobile execution constraints. The study examines hallucination, jailbreak, and noise attacks, and analyzes how platform differences affect model robustness.

58 models evaluated27 adversarial templates15 threat categories

Paper PDFConference program
View figureHide figure
SecuDevSLM workflow connecting model selection, adversarial corpus and template construction, mobile execution, and detection and analysis.
SecuDevSLM evaluation workflow · Figure 1 from the paper.
Citation

Accepted at ICICS 2026; proceedings details are pending.

@inproceedings{li2026small,
  title = {Small Models, Big Variances: Unveiling the Impact of Execution Environments on {SLMs}' Security},
  author = {Li, Zijun and Tu, Tengfei and Lian, Yuxin and Kong, Zicheng and Zhou, Yiling and Liu, Aofan and Li, Wenmin},
  booktitle = {International Conference on Information and Communications Security (ICICS)},
  year = {2026},
  note = {Accepted for publication}
}
AAAI 2026Published · Conference paper

Semantics-Preserving Adversarial Attacks on Event-Driven Stock Prediction Models

Aofan Liu, Haoxuan Li, Hongjian Xing, Yuguo Yin, Zijun Li, and Yiyan Qi

Proceedings of the AAAI Conference on Artificial Intelligence, 40(44), 37645–37653, 2026.

ChameleonAttack combines adaptive optimization in latent space with a semantic translation module to generate fluent adversarial text. The work investigates the robustness of financial language models to attacks that preserve linguistic meaning.

View figureHide figure
Adversarial suffix optimization and semantic translation stages, followed by the final ChameleonAttack framework.
ChameleonAttack framework · Figure 2 from the paper.
PublicationPaper PDFDOI
Citation
@article{liu2026semantics,
  title = {Semantics-Preserving Adversarial Attacks on Event-Driven Stock Prediction Models},
  author = {Liu, Aofan and Li, Haoxuan and Xing, Hongjian and Yin, Yuguo and Li, Zijun and Qi, Yiyan},
  journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year = {2026},
  volume = {40},
  number = {44},
  pages = {37645--37653},
  doi = {10.1609/aaai.v40i44.41099}
}
ICIVC 2026Published · Conference paper

LibSleuth: Semantic-Aware and Lightweight Third-Party Library Detection for Android Software Supply Chain Security

Ying Yang, Chao Zhou, Can Xu, Kai Wang, and Zijun Li (corresponding author)

11th International Conference on Image, Vision and Computing (ICIVC), 873–877, 2026.

Semantic-aware, lightweight detection of third-party libraries for Android software supply chain security.

View figureHide figure
LibSleuth workflow with library preprocessing, core class identification, LLM coding-semantic feature extraction, and multi-round matching.
LibSleuth overview · Figure 1 from the paper.
IEEE XploreDOI
Citation
@inproceedings{yang2026libsleuth,
  title = {{LibSleuth}: Semantic-Aware and Lightweight Third-Party Library Detection for Android Software Supply Chain Security},
  author = {Yang, Ying and Zhou, Chao and Xu, Can and Wang, Kai and Li, Zijun},
  booktitle = {2026 11th International Conference on Image, Vision and Computing (ICIVC)},
  year = {2026},
  pages = {873--877},
  publisher = {IEEE},
  doi = {10.1109/ICIVC70586.2026.11686618}
}
IVMSP 2026Published · Workshop paper

ActiveVideoAgent: Agent-Driven Spatio-Temporal Indexing for Long Video Reasoning

Xi Guo, Lili Jia, Jiawei Zhang, Zijun Li, Jibo Qin, Liye Zhang, Zhensheng Di, and Guan Wang

IEEE 15th Image, Video, and Multidimensional Signal Processing Workshop (IVMSP), 1–5, 2026.

Agent-driven spatio-temporal indexing for reasoning over long videos.

View figureHide figure
ActiveVideoAgent framework showing video preprocessing, multimodal knowledge indexing, and retrieval and generation with a multi-agent reasoning core.
ActiveVideoAgent framework · Figure 2 from the paper.
IEEE XploreDOI
Citation
@inproceedings{guo2026activevideoagent,
  title = {{ActiveVideoAgent}: Agent-Driven Spatio-Temporal Indexing for Long Video Reasoning},
  author = {Guo, Xi and Jia, Lili and Zhang, Jiawei and Li, Zijun and Qin, Jibo and Zhang, Liye and Di, Zhensheng and Wang, Guan},
  booktitle = {2026 IEEE 15th Image, Video, and Multidimensional Signal Processing Workshop (IVMSP)},
  year = {2026},
  pages = {1--5},
  publisher = {IEEE},
  doi = {10.1109/IVMSP68335.2026.11618820}
}
arXiv 2025Preprint

TMCIR: Token Merge Benefits Composed Image Retrieval

Chaoyang Wang, Zeyu Zhang, Long Teng, Zijun Li, and Shichao Kan

Composed image retrieval combines a reference image with a text description of the desired change. TMCIR studies intent-aware cross-modal alignment and adaptive token fusion to balance visual and textual information.

View figureHide figure
Comparison of visual-dominant, text-dominant, and TMCIR adaptive token merge methods for composed image retrieval, illustrated with a reference image and a desired textual change.
Adaptive token merge for composed image retrieval · Figure 1 from TMCIR.
arXivPaper PDF
Citation
@misc{wang2025tmcir,
  title = {{TMCIR}: Token Merge Benefits Composed Image Retrieval},
  author = {Wang, Chaoyang and Zhang, Zeyu and Teng, Long and Li, Zijun and Kan, Shichao},
  year = {2025},
  eprint = {2504.10995},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  doi = {10.48550/arXiv.2504.10995}
}
2024Journal article

Software Vulnerability Detection Method Based on Abstract Syntax Tree Feature Migration (AST-FMVD)

Zi-jun Li, Tao Li, Hao-dong Chen, Qin Yu, Meng-qing Qiao, Lin Li, Jie Wang, Zhen-hua Wan, and Jing-han Song

Computer Technology and Development, 34(6), 81–88, June 2024.

AST-FMVD maps abstract syntax tree features between programming languages using semantic similarity and context-aware mappings. A Java-trained vulnerability detector is applied to Python code, exploring feature transfer without labeled target-domain training samples.

View figureHide figure
AST-FMVD workflow from target-language source code and abstract syntax trees through feature mapping, vectorization, and vulnerability detection.
AST-FMVD framework · Figure 1 from the paper.
Paper PDFDOI
Citation
@article{li2024astfmvd,
  title = {Software Vulnerability Detection Method Based on Abstract Syntax Tree Feature Migration ({AST-FMVD})},
  author = {Li, Zi-jun and Li, Tao and Chen, Hao-dong and Yu, Qin and Qiao, Meng-qing and Li, Lin and Wang, Jie and Wan, Zhen-hua and Song, Jing-han},
  journal = {Computer Technology and Development},
  year = {2024},
  volume = {34},
  number = {6},
  pages = {81--88},
  doi = {10.20165/j.cnki.ISSN1673-629X.2024.0065},
  language = {Chinese}
}

Patents

Chinese invention patentGranted · 22 September 2026

Zero-Shot Vulnerability Detection Method Based on Transfer Learning

Zijun Li (first inventor), Tao Li, Hao-dong Chen, Qin Yu, Meng-qing Qiao, Lin Li, Jie Wang, Zhen-hua Wan, and Jing-han Song

Patent No. ZL 2023 1 1092273.9 · Grant publication No. CN 117094005 B

Filed 28 August 2023 · Patent holder: Wuhan University of Science and Technology

CNKICNIPA lookupGoogle Patents