[Ppm] [PPM] [CFRG] VDAF to Support Federated Machine Learning

[Junye Chen <[e-mail]>]

Hi PPM & CFRG,

We recently proposed a new VDAF [1]: Private Inexpensive Norm Enforcement (PINE) [2] to support federated machine learning use cases. The idea of federated learning is to train machine learning models with gradients (i.e. model updates) from client devices, without the raw client data leaving the devices. PINE is based on a recent paper [3].

PINE VDAF can be used to verify and aggregate client gradients, expressed as vectors of floating point numbers, with bounded “L2-norm”, which is computed by taking the square root of the sum of squares of all vector entries. Its design is based largely on Prio3 [4] in that the norm is computed and verified using an FLP, but it uses a new technique for verifying the correctness of the norm computation that is incompatible with Prio3. The communication cost and computation cost of PINE should be much better compared to those of Prio3 with a new validity circuit.

We will spend some time discussing this new VDAF in CFRG in IETF119, and also its application on DAP [5] in the next PPM meeting as well. In the meantime, feel free to take a look at the draft and the draft repository [6] which has a reference implementation. Thanks!

Best,
Junye

[1]: https://datatracker.ietf.org/doc/draft-irtf-cfrg-vdaf/08/
[2]: https://datatracker.ietf.org/doc/draft-chen-cfrg-vdaf-pine/00/
[3]: https://arxiv.org/abs/2311.10237
[4]: https://www.ietf.org/archive/id/draft-irtf-cfrg-vdaf-08.html#name-prio3
[5]: https://datatracker.ietf.org/doc/draft-ietf-ppm-dap/10/
[6]: https://github.com/junyechen1996/draft-chen-cfrg-vdaf-pine