Papers · Proving training · VeriML

VeriML

VeriML: Enabling Integrity Assurances and Fair Payments for Machine Learning as a Service

Venue
IEEE TPDS
Date
2021-10-01
Proof system
Commit to iteration inputs/outputs; prove randomly retrieved iterations on demand
Hardware
Intel Core i5-4460S
Numbers from
survey
Paper
https://arxiv.org/abs/1909.06961

Reported benchmarks

modelbatch sizeproving time sverification time s
Neural network32120.99

Notes

Earliest verifiable-training entry. Does not prove every iteration -- pre-commits to several and lets the verifier challenge a subset. Supports linear/logistic regression, NNs, SVM, K-Means, decision trees.

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

Sampling, trading soundness for milliseconds, on purposeThe hardest claim in the tableEvery relaxation, and what an adversary can still doThe systems, grouped by what they gave up