Source · Tier A · Peer-reviewedSource · Optimistic Verifiable Training by Controlling Hardware Nondeterminism
Optimistic Verifiable Training by Controlling Hardware Nondeterminism
M. Srivastava, S. Arora, D. Boneh. 2024. Advances in Neural Information Processing Systems 37 (NeurIPS 2024).
| Original | https://arxiv.org/abs/2403.09603 |
|---|---|
| arXiv | 2403.09603 |
| Version | arXiv v3 (25 November 2024) read in HTML; v1 posted 14 March 2024. Venue from the arXiv comments field and the code README. The NeurIPS proceedings page was not read. |
| Accessed | 2026-09-25 |
| Note | Code at https://github.com/meghabyte/verifiable-training (not pinned). Exact FP32 replication of ResNet-50 (CIFAR-10) training and GPT-2 fine-tuning across NVIDIA A40, Titan XP and RTX 2080 Ti; trainer overhead 1.2–1.4×, auditor 1.3–1.7×; assumes one of n auditors is honest. |