# Track A - Federated Learning Setting

**URL:** <https://community.drivendata.org/t/track-a-federated-learning-setting/8052>\
**Category:** PETs Prize Challenge\
**Created:** [September 15, 2022, 1:47pm UTC](https://community.drivendata.org/t/track-a-federated-learning-setting/8052 "2022-09-15T13:47:43Z")\
**Posts on this page:** 6\
**Page:** 1

<div class="post-metadata">

**Author:** ![khangtran97](https://avatars.discourse-cdn.com/v4/letter/k/22d042/32.png) [@khangtran97](https://community.drivendata.org/u/khangtran97)\
**Post date:** [September 15, 2022, 1:47pm UTC](https://community.drivendata.org/t/track-a-federated-learning-setting/8052/1 "2022-09-15T13:47:43Z")

</div>

Is it true that the SWIFT knows everything that each bank knows except the flag of the accounts? If it is the case, why does the SWIFT need the banks? SWIFT can train a centralized model and use it without sharing the model with the banks. If so, there would be no privacy risk at all.

---

<div class="post-metadata">

**Author:** ![jayqi](https://yyz2.discourse-cdn.com/flex028/user_avatar/community.drivendata.org/jayqi/32/924_2.png) [@jayqi](https://community.drivendata.org/u/jayqi)\
**Post date:** [September 15, 2022, 2:53pm UTC](https://community.drivendata.org/t/track-a-federated-learning-setting/8052/2 "2022-09-15T14:53:31Z")

</div>

Hi @khangtran97. There is information that the banks have that are relevant signals for the anomaly detection. A model that only has access to SWIFT’s data (which, based on how this use case is structured, is _not_ a _centralized_ model), does not have complete access to all relevant signals, and accordingly is expected to not be as accurate.

---

<div class="post-metadata">

**Author:** ![khangtran97](https://avatars.discourse-cdn.com/v4/letter/k/22d042/32.png) [@khangtran97](https://community.drivendata.org/u/khangtran97)\
**Post date:** [September 15, 2022, 3:29pm UTC](https://community.drivendata.org/t/track-a-federated-learning-setting/8052/3 "2022-09-15T15:29:28Z")

</div>

Thank you for you response. However, the question is still there. Is it true that the SWIFT knows everything that each bank knows except the flag of the accounts?

---

<div class="post-metadata">

**Author:** ![jayqi](https://yyz2.discourse-cdn.com/flex028/user_avatar/community.drivendata.org/jayqi/32/924_2.png) [@jayqi](https://community.drivendata.org/u/jayqi)\
**Post date:** [September 15, 2022, 3:55pm UTC](https://community.drivendata.org/t/track-a-federated-learning-setting/8052/4 "2022-09-15T15:55:26Z")

</div>

@khangtran97

Our role as organizers is to provide conceptual guidance for you to have the necessary information to solve the task. Conceptually, you are have a correct understand that, with the exception of account flags as you have noted, the account information held in both the SWIFT and bank datasets are the same type of account information.

For empirical conclusions regarding the dataset, participants are responsible for analyzing the data and for testing hypotheses that you may have about how to best design a model for solving the problem.

---

<div class="post-metadata">

**Author:** ![khangtran97](https://avatars.discourse-cdn.com/v4/letter/k/22d042/32.png) [@khangtran97](https://community.drivendata.org/u/khangtran97)\
**Post date:** [September 15, 2022, 5:06pm UTC](https://community.drivendata.org/t/track-a-federated-learning-setting/8052/5 "2022-09-15T17:06:47Z")

</div>

So there will be no Federated Learning at all ? Why don’t the banks send the flags, which are not PII, to SWIFT. Then, SWIFT can train a model with complete information. Therefore, we do not need federated learning in this competition.

---

<div class="post-metadata">

**Author:** ![jayqi](https://yyz2.discourse-cdn.com/flex028/user_avatar/community.drivendata.org/jayqi/32/924_2.png) [@jayqi](https://community.drivendata.org/u/jayqi)\
**Post date:** [September 15, 2022, 6:56pm UTC](https://community.drivendata.org/t/track-a-federated-learning-setting/8052/6 "2022-09-15T18:56:53Z")

</div>

Hi @khangtran97

> flags, which are not PII

This is not correct. All data is considered part of the scope of sensitive data for this challenge.

In general, there is information that the bank holds that is relevant to the anomaly detection task that SWIFT does not have access to alone. It is up to participants to figure out how to model this, and how to incorporate that into a privacy-preserving solution using federated learning. A solution that partially solves the modeling problem by not incorporating federated learning does not meet the objectives of the challenge.