Article Details

Title · Authors · Abstract · Keywords · DOI · References

A Split-Inference Large Language Model Privacy-inference Framework for Code Review

LIANG Ying, LIU Chong, WANG Chao

(AVIC Xi'an Aeronautics Computing Technique Research Institute, Xi'an 710068, China)

Section：Network and Information Security · Vol. 2026, Issue 7 · Pages：166-170,175 · 2096-4706（2026）07-0166-06 · CLC number：TP309 · Document code：A · DOI：10.19850/j.cnki.2096-4706.2026.07.032

Abstract

This paper proposes an engineering method based on “Split Inference” to address the potential risk of source code leakage when enterprises invoke the Large Language Model (LLM) in code review scenarios. This method achieves privacy protection for sensitive source code by locally executing tokenizer, embedding, and the initial layers of computation, while running the remaining layers in a remote trusted execution environment. The paper establishes a complete system architecture, analyzes performance and bandwidth requirements, designs a remote attestation process, and introduces multiple security enhancement measures. Through theoretical derivation and experimental design, it validates the trade-off between privacy protection and performance feasibility. Ultimately, this study provides a feasible pathway for deploying large models in security-sensitive domains.

Keywords

Large Language Model; Split-Inference; source code review; privacy protection; remote proof; confidential computing

References

  1. [1] MUDVARI A，VAINIO A，OFEIDIS I，et al. Adaptive Compression-aware Split Learning and Inference for Enhanced Network Efficiency [J/OL].ACM Transactions on Internet Technology，2024，24（4）：1-26（2024-11-15）[2025-10-09].https://doi.org/10.1145/3687471.
  2. [2] ASSINE J S，FILHO J C S S，VALLE E. Single-training Collaborative Object Detectors Adaptive to Bandwidth and Computation [J/OL].arXiv:2105.00591 [cs.CV].（2021-05-03）.https://arxiv.org/abs/2105.00591.
  3. [3] YAO D，XIANG L，XU H，YE H，et al. Privacy-preserving Split Learning via Patch Shuffling over Transformers [C]//IEEE International Conference on Data Mining.Dallas：IEEE，2022：638-647.
  4. [4] SEYED M，YANG L，YAN F，et al. Communication-Efficient Training Workload Balancing for Decentralized Multi-Agent Learning [C]//IEEE International Conference on Distributed Computing Systems.Jersey City：IEEE，2024：680-691.
  5. [5] CHEN P，YANG J R，LIN J X，et al. A Practical Clean-label Backdoor Attack with Limited Information in Vertical Federated Learning [C]//IEEE International Conference on Data Mining.Shanghai：IEEE，2023：41-45.
  6. [6] ERIC S，ANTONIO M，BRAUN T. DISNET: Distributed Micro-split Deep Learning in Heterogeneous Dynamic IoT [J].IEEE Internet of Things Journal，2024：6199-6216.
  7. [7] CHANDRA T，PATHUM C，SEYIT C，et al. Splitfed: When Federated Learning Meets Split Learning [C]//Proceedings of the AAAI Conference on Artificial Intelligence.Vancouver：AAAI，2022：8485-8493.
  8. [8] ANDERS N. Artificial Intelligence and Climate Change: Ethical Issues [J/OL].Journal of Information, Communication and Ethics in Society，2023，21（1）：1-15（2023-01-31）[2025-10-11].https://doi.org/10.1108/JICES-11-2021-0106.
  9. [9] WADHWA M，GUPTA G R，SAHU A，et al. Pfsl: Personalized & Fair Split Learning With Data & Label Privacy for Thin Clients [C]//IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing.Bangalore：IEEE，2023：377-390.
  10. [10] Qin Zhen, Zhuang Tianming, Zhu Guosong, etc Overview of Security Attacks and Defense Strategies for Artificial Intelligence Models [J]. Computer Research and Development, 2024, 61 (10): 2627-2648
  11. [11] LIU H M，JIN F，ZENG H，et al. Image Enhancement Guided Object Detection in Visually Degraded Scenes [J].IEEE Transactions on Neural Networks and Learning Systems，2024，35（10）：14164-14177.

About the authors

Liang Ying (1981.01-), female, Han ethnicity, from Xunyang, Shaanxi, is a project manager with a master's degree. Her research interests include artificial intelligence, big model applications, software development project management, and agile development.