Source · Tier A · Peer-reviewedSource · zkDL: Efficient Zero-Knowledge Proofs of Deep Learning Training
zkDL: Efficient Zero-Knowledge Proofs of Deep Learning Training
H. Sun, T. Bai, J. Li, H. Zhang. 2025. IEEE Transactions on Information Forensics and Security 20, pp. 914–927.
| Original | https://arxiv.org/abs/2307.16273 |
|---|---|
| DOI | 10.1109/TIFS.2024.3520863 |
| arXiv | 2307.16273 |
| Version | Read arXiv v2 (5 December 2023) on 2026-10-17: the abstract, §I (introduction and threat model) and §VI (experiments, Table I). Journal venue, volume, pages and DOI from the Crossref record; the journal version was not read. Table I averages costs per training step. Its sub-second proving times hold when one proof aggregates 16 to 1,024 training steps, for an 8-hidden-layer ReLU perceptron with 1,024 neurons per layer (about 10 million parameters) on CIFAR-10, on an A100 GPU. With aggregation only within each step, proving took about 6 seconds per step. |
| Accessed | 2026-10-17 |