1) I've reviewed the "Evaluation of Cohort Algorithms for the FLoC API" @ github.com/google/ads-pri… & have thoughts..🧡

high-level takeaway is that both methods Google tested *require an anonymity server* to filter cohorts that are too small.

This is *not* a deal breaker* Image
2) *Google tested methods that required an anonymity server because they don't have federated learning built into Chrome.

So Google tested "Centralized cohort building/filtering" vs "Pseudo-on-device cohort building/filtering" - the privacy safe version was 85% of the quality.
3) Differential privacy β‰  K-anonymity / We should focus on K-score to protect users (& merge cohorts) - it's a subtle difference but K-scores are more easily integrated into a "minimum viable cohort size" to be built into an open source anonymity server or federated Learning.. ImageImage
4) Google can't be in charge of the anonymity server UNLESS it's also open source + audited by any/all DSP partners / gov's who request access.

5) Google should plan to not be trusted, and that seems to be strongly how they are approaching it w/ federated learning suggestions.. Image
6) imo Apple is the defacto leader in federated learning privacy architecture -- tons of papers from their team like machinelearning.apple.com/research/prote… & machinelearning.apple.com/research/learn… to better understand how the raw browsing history parsing MUST be federated if these domain-based hashes are used
7) Google used Chrome ingestions to compare a Fully Centralized Version (which should* NEVER be deployed) to a Decentralized Version (yay) -- this comparative process is essential & I have no concerns w/ testing it.

**Decentralized was 85% of the quality of centralized** (yay!) ImageImage
8) I'm a fan of google explaining some of the options for creating Cohorts -- there is a MAGIC "middle ground" where too small of cohorts will eventually be filtered out (aka, not enough traffic on your professional bowling news pages, so no cohorts womp womp) - big topics rule~ ImageImage
9) Publishers will create their own cohorts w/ k-anon filters for too small cohorts. This will require very interesting traffic analysis to find the ~10 biggest categories across a website - but if you have million of visitors per month, you could do 20-30-40 cohort categories.. Image
10) Topic clustering proved to Google that there are fewer ways to "be wrong" when it comes to cohort/subcohort filtering *(which is very good). Publishers will likely create 50 cohorts, some won't normally be big enough to not be filtered, but rolled-up sub->Pcat still has value Image
11) Ad tech tried centralized servers (adidentity.org) - this was a huge effort that failed even with the seemingly-valid idea to have multiple "processing servers" that orgs could chose from.

imo privacy sandboxes that do NOT have federated learning = privacy failure
12) Federated learning would hash the user's browsing history by domain etc - but then passing these hashes to the anonymity server is where this could also fail if the server isn't open source & cohort filtering isn't agreed upon to be working to protect users (i think it can)
13) My final thought on the Google Privacy Sandbox proposal and cohort creation process is that this process continues to align w/ my proposal for a cohort-segment-basket stock market, due to the fact that these recipes & filters will need to be public. medium.com/@thezedwards/t… πŸ––

β€’ β€’ β€’

Missing some Tweet in this thread? You can try to force a refresh

Keep Current with β„¨π”žπ” π”₯ π”ˆπ”‘π”΄π”žπ”―π”‘π”°

β„¨π”žπ” π”₯ π”ˆπ”‘π”΄π”žπ”―π”‘π”° Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!


Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @thezedwards

29 Oct 20
This is some of the worst ad tech research I’ve ever seen. The markup doesn’t have access to the actual bidding details of either campaign - they don’t have exclusion data either.

A few FB buying facts:

1) Exclusion audiences save money when high-bid pages are in an audience.
2) custom audiences cost less than native FB targeting of page interests/likes

3) lookalikes cost less than custom audiences, and less than native FB targeting

4) campaigns bid against each other - hugely popular states like Florida has tons of competition
5) it’s possible to attack the CPM rates by buying ads against XYZ fan page. Take 40 ads accounts you control, bid on only fan pages (Obama/Biden,hrc) & bid very high. Biden’s optimization choices for a campaign could then be used to push his CPM rates in some markets sky-high.
Read 8 tweets
19 Jul 20
I think there is another big twitter hack going on, but not to verified accounts... rayban DM spam going really big
Here's my threads w/ screen shots of users complaining about this recent twitter hack / DM mass messages pushing RayBans... (Seems like rate limits are being stomped imo)

Read 5 tweets
10 Feb 20
There is a data supply thread going around with experts dunking on people who believe microphones are listening to them & using their conversations to show targeted ads.

Experts claim it’s impossible, but they are wrong...and numerous companies have tech to do it. Thread time..
Numerous companies have been testing this Speech->Keyword->Ad Segments for over a decade. I’m sharing 14 patents from major companies that do *exactly this*

Sony submitted for a patent for Speech to Keyword to Advertising tech in 2007 (received in 2011) @ patents.google.com/patent/US20080…
AT&T submitted for a patent for Speech to Keyword to Advertising tech in 2008 (received in 2015) @ patents.google.com/patent/US90150…
Read 24 tweets
1 Nov 19
Just a little reminder that Fitbit purchased TwineHealth.com in 2018 --- Twine was the HIPPA Custom Audience tool used by Fitbit + ties them into insurance/doctor/phRma networks

Twine Audience / Presentations are being purged @ webcache.googleusercontent.com/search?q=cache…
"Helping patients become the masters" - a presentation from Google's new acquisition - the FitBit + Twine Patient Data connection!

Twine video @
This participant upload process used by Fitbit's Twine (now owned by Google) certainly seems like Fitbit has a bunch of patient data that is not HIPPA compliant and should never have been appended to the wearable data.
Read 8 tweets
18 Sep 19
yuck to the plaintext permutive custom params with client keyword blacklists clearly visible in page requests -- this is from the @Independent - here's one gross section:

@Independent * just a reminder -- the UK Independent is a NEWS website, yet they have advertising blocks for Yemen + Saudi Arabia + Houthis etc // this means that those articles do not monetize properly, aka there is a disincentive to cover the war due to advertisers being allowed to block.
**important clarification β€”- permutive is the first Params but ***did not create the keyword blocklists**** I have confirmed from the permutive team this wasn’t theirs but is injected by another partner in the ecosystem. Will follow up

Thanks p-team
Read 8 tweets
5 Sep 19

In May I was commissioned by Brave to help their team audit the Google auctions so they could file details in their GDPR complaint. My pixel auditing background is significant and I’ve built enterprise analytics for nearly 8 years @victorymedium
Here's one of the legacy threads i've done on pixel shenanigans & overlapping liabilities in the online advertising ecosystem due to how Javascript pixels piggyback from a 3rd party to 4th/5th/6th parties. It creates companies reliant on each other.

Before we begin, I keep VERY close tabs on Google systems. I'm a Google Certified Partner, I've taken and passed the Google Analytics certification, and i've been using tools like Google Tag Manager for over 7 years (since basically the start). I'm deep into Google pixels...
Read 20 tweets

Did Thread Reader help you today?

Support us! We are indie developers!

This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Too expensive? Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal Become our Patreon

Thank you for your support!

Follow Us on Twitter!