Source · Tier B · PreprintSource · VeriLoRA: Fine-Tuning Large Language Models with Verifiable Security via Zero-Knowledge Proofs
VeriLoRA: Fine-Tuning Large Language Models with Verifiable Security via Zero-Knowledge Proofs
G. Liao, T. Wang, S. Zhang, J. Zhang, L. Shi, D. Tao. 2025. arXiv.
| Original | https://arxiv.org/abs/2508.21393 |
|---|---|
| DOI | 10.48550/arXiv.2508.21393 |
| arXiv | 2508.21393 |
| Version | First posted 29 August 2025; arXiv v3 dated 2 December 2025. The abstract and HTML, read on 2026-09-25, use the title VeriLoRA and state that the paper was accepted for publication at NDSS 2026. |
| Accessed | 2026-09-25 |
| Note | Proves one LoRA fine-tuning iteration (forward pass, backward pass, parameter update) on a single-sample minibatch for LLaMA-3.2 3B and 11B, LLaMA-2 7B and 13B and OPT 6.7B and 13B on one NVIDIA A100 80 GB GPU. The full text read reports 121.93–249.38 s of proving, 156–554 s of commitment generation and 1.87–3.73 s of verification per step. Code linked from the paper at https://github.com/liaoguofu/zkLoRA (MIT licence, README titled VeriLoRA, built on the zkLLM code base); commit not pinned. |