zkCNN

zkCNN: Zero Knowledge Proofs for Convolutional Neural Network Predictions and Accuracy

Tianyi Liu, Xiang Xie, Yupeng Zhang

Venue
ACM CCS 2021
Date
2021-05-25
Proof system
GKR + sum-check (with a linear-time sumcheck for FFT/convolution)
Hardware
AMD EPYC 7R32, 128 GB RAM
Numbers from
primary
Paper
https://eprint.iacr.org/2021/673
Quantization
not stated

Reported benchmarks

modelparamslayersproving time sproof size mbverification time s
VGG1615M1688.30.3410.0593

Notes

Foundational. Linear-prover-time sumcheck for 2D convolution (asymptotically faster than computing the convolution directly) and an O(N) sumcheck for FFT. 1264x faster than prior schemes. Also proves accuracy over a public dataset (20 images). STILL THE FASTEST PROVER ON SMALL MODELS, FIVE YEARS ON. Bionetta (Table 4) re-measures zkCNN on a 16-thread Xeon and finds it proving LeNet5 in 1.05 s and VGG11-mini in 4.10 s -- beating Bionetta's own 3.75 s and 7.70 s. Bionetta only overtakes it above roughly 2M parameters (its Fig. 6), and cannot beat it below that. What zkCNN pays for that prover is the thing GKR always pays: proofs of 23-43 kB against Bionetta's 0.88 kB, and verification of 0.3-1.45 s against 10-20 ms. It also cannot run ResNet18 or MobileNetV2 at all in Bionetta's harness (unsupported operators). The trade has not changed since 2021; only the models have.

Quantization, in full

bits
None
scheme
Affine: a = L(q - Z), with scale L and zero-point Z. The scheme zkGPT later inherits.
accuracy_retention
None

Our reading

Citation neighbourhood

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

Discussed in

Other recorded fields
objectives:
- inference
- testing