1462109 - Federated Learning client

Status: RESOLVED FIXED

(Data Platform and Tools :: General, enhancement, P1)

Description

[Florian Hartmann]

To train a Federated Learning model, clients need to periodically talk to a server and report their proposed model updates.
For our application, this client part needs to work in Firefox.

Comment 1

[Florian Hartmann]

We decided to implement the client-side part using an addon.
A first minimal working version of it can be found on GitHub [1].

It was implemented as a WebExtension.
Because some legacy addon APIs are necessary to make it work, there are some experiment APIs in the WebExtension.

The model tries to predict what history suggestion users are going to click on.
To do this the frecency algorithm [2] is used.
Federated learning is going to optimize the weights used in that algorithm.

[1] https://github.com/florian/federated-learning-addon
[2] https://developer.mozilla.org/en-US/docs/Mozilla/Tech/Places/Frecency_algorithm

Comment 2

[Florian Hartmann]

The draft of the addon is now fully functional.
Shield study related parts and Telemetry was added as well.
The first version is now being reviewed on GitHub [1].

[1] https://github.com/florian/federated-learning-addon/pull/1

Comment 3

[Florian Hartmann]

Attachment: data_review.txt

Comment 4

[Chris H-C :chutten]

Comment on attachment 8988354 [details]
data_review.txt

DATA COLLECTION REVIEW RESPONSE:

    Is there or will there be documentation that describes the schema for the ultimate data set available publicly, complete and accurate?

Yes.

    Is there a control mechanism that allows the user to turn the data collection on and off?

Yes. Standard Telemetry mechanisms apply.

    If the request is for permanent data collection, is there someone who will monitor the data over time?

N/A, time-bound study.

    Using the category system of data types on the Mozilla wiki, what collection type of data do the requested measurements fall under? **

Category 2.

    Is the data collection request for default-on or default-off?

Default-on.

    Does the instrumentation include the addition of any new identifiers (whether anonymous or otherwise; e.g., username, random IDs, etc. See the appendix for more details)?

No.

    Is the data collection covered by the existing Firefox privacy notice? 

Yes.

    Does there need to be a check-in in the future to determine whether to renew the data?

No, the time bound of the study is firm.

---
Result: datareivew+

Comment 5

[Florian Hartmann]

Attachment: federated_learning_awesome_bar-1.0.zip

This is the addon that needs to be signed.

Comment 6

[Rehan Dalal (Inactive)]

Attachment: [e-mail]

Comment 7

[Florian Hartmann]

Attachment: federated_learning_awesome_bar-1.0.zip

This is the addon that needs to be signed.
It now has a @shield.mozilla.org ID.

Comment 8

[Rehan Dalal (Inactive)]

Attachment: [e-mail]

Comment 9

[Florian Hartmann]

The code review was completed and we submitted the addon to QA now.

Comment 10

[Florian Hartmann]

Attachment: federated_learning_awesome_bar-1.0.zip

QA asked for an easier way to change the study variation, so this build adds that functionality.

Comment 11

[Florian Hartmann]

Attachment: federated_learning_awesome_bar-1.0.zip

QA asked for an easier way to change the study variation, so this build adds that functionality.

Comment 12

[Rehan Dalal (Inactive)]

Attachment: [e-mail]

Comment 13

[Florian Hartmann]

Attachment: federated_learning_awesome_bar-1.0.xpi

This version of the addon fixes two bugs that QA found [1, 2].

[1] https://github.com/florian/federated-learning-addon/issues/7
[2] https://github.com/florian/federated-learning-addon/issues/9

Comment 14

[Rehan Dalal (Inactive)]

Attachment: [e-mail]

Comment 15

[Florian Hartmann]

Attachment: federated_learning_awesome_bar-1.0.0.xpi

This version properly resets the changed prefs when it is uninstalled or disabled.

Comment 16

[Michael Cooper [:mythmon]]

Attachment: [e-mail]

Comment 17

[Florian Hartmann]

The addon was deployed to users two days ago.
So far it seems like everything is working nicely.
There could be future improvements though, like a better way of fetching a model after the user just got an internet connection again.