What is Google FloC? Federated Learning of Cohorts Explained

This article explains Federated Learning of Cohorts (FLoC or “flock”)  and how it can power interest-based advertising on the web without cookies while maintaining user privacy.

The FLoC proposal is part of Google's Privacy Sandbox, an initiative to build a privacy-by-design browser crafted to sustain an ad-supported web similar to the one we know today.

Eliminating the third-party cookie is a key requirement of this initiative but will completely change how previously established mechanisms of advertising, like interest-based advertising, work.

Tracking users around the web via a third-party cookie is the basis of behavioral and interest-based advertising, but is not respectful of user privacy due to its opaque and passive nature. Flocks could reduce advertiser reliance on cookies for interest-based targeting.

Google also released a separate proposal called TURTLEDOVE, to power behavioral advertising. Follow the link to read the explainer on that proposal.

What is a flock?

Federated learning is a machine learning technique that allows many clients (browsers) to work together to form a centralized model without exchanging actual sample data. This is the same method used by predictive mobile keyboards to improve word suggestions without sending actual text information off a device.

Google proposes for browsers to use federated learning to create cohorts or “flocks” of users with similar interests based on the URLs visited or the content of those pages. The collected data would stay local and never leave the client device, preserving user privacy.

The browser would assign a label to each flock it creates such as “52P9”, which would be accessible via an HTTP header.

GET https://adpartner.com/serve_ad.html?width=300&height=250
Referer: https://videogamenews.com/home.html
Sec-CH-Flock: 52P9

The browser would update the flock over time based on a user’s browsing behavior.

Google suggests that flocks should consist of thousands of users to bolster anonymity. They also suggest short four-digit flock labels to make clear that there are a low number of total flocks, reinforcing the general nature of each flock.

Flocks and Advertising

Ad tech platforms could ingest a flock label with every ad request to accomplish specific goals.

Advertisers could record flock labels every time a user checks out on their site and categorize the labels by product interest. They could create flocks based on the items added to a cart or actually purchased.

An online art shop could note that “39X1” flock members tend to purchase nature photography more than any other flock. The art shop could then use an ad network to target this particular flock with creatives that feature nature photography prints.

Ad tech platforms could also ingest flock labels as input parameters in their own models, developed to optimize towards a particular goal.

A model could optimize for clicks or view-through-rate by observing that certain flocks result in more clicks on an ad, or other flocks tend to view a video ad all the way to the end. The model could then tweak bidding behavior to favor these flocks based on an advertiser’s goals.

Benefits of Flocks

Flocks would eliminate the need for advertisers to drop third-party cookies on a user’s browser and track them across the web to target them by interest — a huge win for user privacy.

Advertisers use Interest-based advertising to serve ads that they believe will be relevant to a user.

For example, someone who spends most of their time browsing video game news would not be interested in seeing ads for scuba diving vacations. Serving ads not relevant to a user results in wasted ad spend from displaying ads to the wrong audience.

Flock Criticisms

Google admits that one flaw of Federated Learning of Cohorts is that a browser may inadvertently use sensitive categories as part of its data collection.

A browser could potentially block certain sensitive data, but identifying which categories are sensitive could be challenging and change from user to user. A category may not be sensitive to one user but could be very sensitive to another.

Other criticisms focus on the idea that any method of targeting by interest is unacceptable. Even though flocks do not require advertisers to hold your full browsing history, some privacy advocates point out that they don’t want their habits or interests targetable in any way.

The Electronic Frontier Foundation was quick to paint a sinister picture of flocks:

The degree to which this statement is true will all depend on implementation. Browsers would have to ensure that flocks are well distributed and general enough to satisfy privacy advocates like the EFF. If a flock is too specific or does not properly ignore sensitive categories, the privacy benefit of flocks would erode quickly.