Google Git
Sign in
android / platform / external / federated-compute / 6ed21516de1c58f36dd17a87f34709f976e1e01e
commit6ed21516de1c58f36dd17a87f34709f976e1e01e[log] [tgz]
authorXin Li <[email protected]>Sat Sep 07 02:04:37 2024 +0000
committerAutomerger Merge Worker <[email protected]>Sat Sep 07 02:04:37 2024 +0000
treee298dcd669ed5654973f267d7c9699106bb73772
parentbaf14b4c828d9d08851bce47c5add99e50cf909c [diff]
parent5e173a7baa328e844eac1516a566f50215851386 [diff]
[automerger skipped] Merge 24Q3 to AOSP main am: 5e173a7baa -s ours

am skip reason: Merged-In Id3eb609bcb1c60088b07dfe1a5c059d6b4354fe5 with SHA-1 1591b1702a is already in history

Original change: https://android-review.googlesource.com/c/platform/external/federated-compute/+/3258592

Change-Id: Ia858b3432586aeb98aad606fd475696aa3fecd85
Signed-off-by: Automerger Merge Worker <[email protected]>
tree: e298dcd669ed5654973f267d7c9699106bb73772
  1. fcp/
  2. third_party/
  3. Android.bp
  4. CONTRIBUTING
  5. GETTING_STARTED.md
  6. LICENSE
  7. METADATA
  8. MODULE_LICENSE_APACHE2
  9. OWNERS
  10. PREUPLOAD.cfg
  11. README.md
  12. requirements.txt
  13. TEST_MAPPING
  14. WORKSPACE
README.md

Federated Compute Platform

This repository hosts code for executing federated programs and computations.

Definitions

A federated computation is a set of processing steps that run on a server and set of clients, where each step is either

  • local processing of values on the client or server or
  • transport which moves values via
    • broadcast (server-to-clients)
    • select (client-server response)
    • aggregate (clients-to-server). A federated computation invoked by a central coordinator returns one or more aggregates.

A federated program is a set of processing steps run by a central coordinator that include one or more invocations of federated computation, transformations of computation results, and releases of aggregate values to the engineer/analyst who invoked the program.

To learn more about these concepts, check out:

Infrastructure

At Google, federated programs and computations are authored in TensorFlow Federated, compiled to deployable artifacts, and run in a distributed system consisting of a central coordinator, and a set of devices such as phones. The TFF repository contains infrastructure for authoring and simulating federated programs and computations.

This repository hosts infrastructure for compiling and running federated programs and computations in the cross-device setting. We are actively working on open sourcing the core components of our production infrastructure, with a focus on privacy-sensitive code-paths such as the pipeline for compiling deployable artifacts from TFF computations, client-side processing, and server-side aggregation logic.

As of 12/7/2022, parts of the repository - in particular, code in the client/ directory, and the service & data format definitions in proto/ - are used in production in Google's federated learning infrastructure. Other parts - notably, production server side infrastructure - have not yet been open sourced due to its dependencies on proprietary infrastructure, and we instead provide a reference / example server implementation in demo/ for demonstration purposes.

The best way to get started is to run the end-to-end demo //fcp/demo:federated_program_test, which will spin up example services, clients, and run a federated program; this test will cover the majority of the code in this repository.

This is not an officially supported Google product.

Getting Started

Please refer to the instructions in GETTING_STARTED.md.