SpaGKR

Sparsity-aware protocol for zk-friendly ML models: Shedding lights on practical ZKML

A. Li, Q. Liang, M. Dongunverified

Venue
IACR ePrint 2024
Date
2024-06-01
Proof system
GKR / sum-check, modular; composes with Lasso
Numbers from
survey
Paper
https://eprint.iacr.org/2024/1018
Quantization
1.58 bits

Reported benchmarks

modelspeedup vs dense zkmlspeedup from ternary
Sparse linear layers45
Ternary network5

Notes

Two orthogonal wins, both worth separating when plotting: (1) SpaGKR-LS makes proof time scale with the number of NON-ZERO parameters, giving 45x on sparse linear layers; (2) ternary networks give a further ~5x. Relevant to MoE and pruned models.

Quantization, in full

bits
1.58
scheme
Ternary networks: parameters restricted to {-1, 0, 1}, which ELIMINATES multiplications entirely
accuracy_retention
None
speedup_from_quantization
5.0
note
THE key data point for the quantization question. Ternary quantization alone buys a further ~5x proof-time reduction, on top of the 45x that sparsity-aware linear layers (SpaGKR-LS) buy over sparsity-oblivious ZKML. This is the strongest published evidence that lower bit width => faster proving, and it quantifies the effect.

Our reading

Citation neighbourhood

Builds on
  • none recorded
Cited by, in this corpus

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

Discussed in