Papers · Proving training · zkDL

zkDL

zkDL: Efficient Zero-Knowledge Proofs of Deep Learning Training

Haochen Sun, Tonghe Bai, Jason Li, Hongyang Zhang

Venue
IEEE TIFS 2024
Date
2023-07-30
Proof system
zkReLU (bespoke proof for ReLU + backprop) over FAC4DNN circuit design
Hardware
12 CPU cores, Tesla A100 GPU, 128 GB RAM; CUDA
Numbers from
primary
Open source
yes
Paper
https://arxiv.org/abs/2307.16273
PDF
https://arxiv.org/pdf/2307.16273
Code
https://github.com/SafeAILab/zkDL

Reported benchmarks

modelparamsbatch sizeproving time sproof size mbverification time s
8-layer DNN10M640.863.3e-050.19

proving_time_note: per batch update; paper says <1s

proof_size_note: survey Table IV: 0.033 KB. Suspiciously small -- unverified.

Notes

FAC4DNN aggregates proofs across layers and training steps without being constrained by their sequential order -- the key trick that makes per-batch proving sub-second. The 33-byte proof size in the survey's Table IV looks implausible; verify against the paper.

Our reading

Citation neighbourhood

Builds on
  • none recorded
Cited by, in this corpus
  • none recorded

Edges are a proxy: paper A's text mentions B anywhere (body or bibliography). See the full graph.

Discussed in

The hybrid cellThe hardest claim in the tableEvery relaxation, and what an adversary can still doThe systems, grouped by what they gave up
Other recorded fields
eprint: https://eprint.iacr.org/2023/1174