Papers · Federated & aggregation · PRoVeFL

PRoVeFL

PRoVeFL: Private Robust and Verifiable Aggregation in Federated Learning

Harsh Kasyap, Anil Kumar Pradhan, Ugur Ilker Atmaca, Graham Cormode, Carsten Maple

Venue
arXiv:2607.06612
Date
2026-07-07
Hardware
Intel Xeon E5-4650 v4 @ 2.20 GHz, 112 CPUs, 256 GB RAM
Numbers from
primary
Uses ZK
no
Paper
https://arxiv.org/abs/2607.06612
PDF
https://arxiv.org/html/2607.06612v1

Reported benchmarks

modelparamsaggregation ruleclientswallclock s
LeNet5 (CIFAR-10 S)62K———
ResNet-18 (CIFAR-10 L)273K———
LSTM (Shakespeare)818K———
——Krum2002000
——Trimmed-Mean200400
——FLTrust20020

params_note: As reported. Standard ResNet-18 is ~11M params, so this must be a reduced variant -- worth checking.

Notes

These wallclock numbers are per-round aggregation cost, NOT proving time. Plotting them against zkPoT proving times would be a category error.

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 cellWhat each system actually proves
Other recorded fields
uses_zk_note: 'VERIFIED: the strings "zero-knowledge", "ZKP", "SNARK" and "proof of
  training" do not appear anywhere in the paper. Do not file this under zkML.'
primitive: Multi-key FHE (Ring-LWE) + discrete-log commitments; bilinear pairwise
  checks
what_is_proven: That a server performed the aggregation rule's arithmetic correctly
  (distances for Krum, differences for Trimmed-Mean, inner products for FLTrust).
  Verified BY ANOTHER SERVER, assuming >=1 server is honest. Not publicly verifiable,
  not succinct, not zero-knowledge.
communication_gb:
- 0.001
- 26