EUVIMEDEUVIMED entdecken
Crossref

PC-FT: A Lightweight and Utility-Preserving Private Fine-Tuning Framework for Edge-Based Agentic AI

Peisheng Zhang, Renao Zhang, Zefang Lv, Qing Fan, Shuai Yuan, Zhitao Guan

IEEE Transactions on Network Science and Engineering · 2026 · 13 · 9648-9660

DOI: 10.1109/tnse.2026.3694751 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Network Science and Engineering, 2026. Verfasst von Peisheng Zhang und 5 weiteren Autor:innen. Die Arbeit nennt 72 Literaturangaben. Gefördert von: Fundamental Research Funds for the Central Universities.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Large language model (LLM)-based agentic AI systems are increasingly moving to the network edge, driven by the need for intelligent services with minimal latency. However, the resource-constrained and data-local nature of edge nodes imposes higher demands on both privacy protection and computational efficiency. Existing differentially private fine-tuning methods often incur high overhead or substantial utility loss. We propose PCFT, a lightweight and utility-preserving differentially private fine-tuning framework designed for LLM-based agentic AI operating at the network edge. PC-FT improves the privacy–utility trade-off through contribution-guided adaptive noise perturbation, enabling parameter-efficient private adaptation. Specifically, it introduces an objective-coupled contribution score derived from a first-order Taylor proxy to quantify each parameter's influence on the training objective. This score partitions parameters into high- and low-contribution groups to enable differentiated noise injection. Furthermore, a lightweight dynamic partial update mechanism is introduced to periodically adjust the partition during training using only already-privatized gradients without incurring additional privacy cost. When benchmarked on E2E and DART, PCFT provides measurable utility improvements over current DP-PEFT baselines without relaxing the privacy guarantee or incurring substantial additional overhead. Privacy attack evaluations confirm that PCFT does not amplify information leakage and often reduces it, thereby achieving a favorable privacy–utility balance for agentic AI deployed at the network edge.

Förderung

  • Fundamental Research Funds for the Central Universities

Zitieren

APA

Zhang, P., Zhang, R., Lv, Z., Fan, Q., Yuan, S., & Guan, Z. (2026). PC-FT: A Lightweight and Utility-Preserving Private Fine-Tuning Framework for Edge-Based Agentic AI. IEEE Transactions on Network Science and Engineering, 13, 9648-9660. https://doi.org/10.1109/tnse.2026.3694751

Vancouver

Zhang P, Zhang R, Lv Z, Fan Q, Yuan S, Guan Z. PC-FT: A Lightweight and Utility-Preserving Private Fine-Tuning Framework for Edge-Based Agentic AI. IEEE Transactions on Network Science and Engineering. 2026;13:9648-9660. doi:10.1109/tnse.2026.3694751

BibTeX
@article{Zhang2026,
  author = {Zhang, Peisheng and Zhang, Renao and Lv, Zefang and Fan, Qing and Yuan, Shuai and Guan, Zhitao},
  title = {PC-FT: A Lightweight and Utility-Preserving Private Fine-Tuning Framework for Edge-Based Agentic AI},
  journal = {IEEE Transactions on Network Science and Engineering},
  year = {2026},
  volume = {13},
  pages = {9648--9660},
  doi = {10.1109/tnse.2026.3694751}
}
Kostenlos starten →