5 ms·

It's actually not given to Google in a user identifiable way. I am not privy to the details, but an overview was presented at a crypto conference: https://www.

It's actually not given to Google in a user identifiable way. I am not privy to the details, but an overview was presented at a crypto conference: https://www.youtube.com/watch?v=ee7oRsDnNNc&t=10m1s https://www.youtube.com/watch?v=ee7oRsDnNNc&t=10m1s It looks to me like a form of homomorphic encryption where both the aggregate consumer transactions (people bought X at merchant M), and the aggregate Google ad information (people saw ad for X at merchant M) are encrypted and the set intersection is computed on the ciphertext revealing only the aggregate number of purchases, not the individuals. Now granted, there could be a security hole in this scheme as with any crypto-protocol, so it does need peer review, but the design, as stated in that YouTube presentation, is not to reveal to Google personal purchasing information, nor reveal to merchants, personal profile information. No personal financial data or purchasing information is supposedly transmitted. The problem with this kind of technology is that it is really difficult to explain to the public, and easily demagogued.

ocdtrekkie 9y ago
My question is... how is that supposed to work? How do they turn "person A's credit card account" and "person A's Google account" into something comparable. To say "X customers saw the ad and then purchased it", you've gotta be able to be able to confirm that the same person is on both sides of the dataset, which is a pretty hard problem even without it allegedly not sharing the info on one side with Google. I guess this complaint from this article is asking the FTC to get more information on this process, or get Google to explain how it works, or something. Upvote for the explanation.
- sloppycee 9y ago
  You can use geography, I assume this is meant to measure ROAS. I've worked on a similar problem, and some very large retailers we worked with wanted to measure ROAS. They turned off all adspend in a certain large metropolitan area, while we turned on in that area. Measuring the sales lift/drop, gave an idea of the platform's effectiveness. Any other demographic data point could also be used, not just geo. You don't necessarily need to link people directly.
- UncleMeat 9y ago
  Homomorphic encryption to do set intersection. With sufficiently large sets, you don't need to do any sort of collection at the individual level.
- singron 9y ago
  I don't know in this case, but there are several identifiers that both parties might have (phone number, email address, billing address, name). In the industry, you can sometimes get "anonymized" data that is keyed by a hash of these identifiers, then you can hash your own identifiers and join to deanonymize. This isn't how it works in this case, but I know other companies do this.
hedora 9y ago
Assuming the encryption works as advertised, these schemes still allow more de-anonymization than you'd think. If you can compute very specific set intersections you can then get the set size low enough to reveal personal information with 100% certainty. (The set of people that bought X at M and Y at N and saw ad A at P and ad B at Q on a particular day might contain one person) I do not know of any work in homomorphic encryption that claims to provide privacy guarantees in the face of side-channels from other sources (in this case, the side-channels could be phone location information for brick-and-mortar purchases, or web browsing activity timestamps for online purchases). It's not surprising that such work is hard to find, since merely computing correct query results breaks most reasonably privacy guarantees. Also, the field of homomorphic encryption encourages people to use horrible cipher-modes like code-book, so guilt-by-association makes me highly suspicious of stuff in that space: https://en.m.wikipedia.org/wiki/Block_cipher_mode_of_operation#Electronic_Codebook_.28ECB.29 https://en.m.wikipedia.org/wiki/Block_cipher_mode_of_operati... The closest thing I know of that is workable is differential privacy, which (in laymans terms) introduces enough noise to make sure that overly-specific queries return random (but representative) garbage. Apple says they use differential privacy to train models on data that's stored on iPhones without actually giving themselves access to that data. Cool stuff.

ocdtrekkie 9y ago

My question is... how is that supposed to work? How do they turn "person A's credit card account" and "person A's Google account" into something comparable. To say "X customers saw the ad and then purchased it", you've gotta be able to be able to confirm that the same person is on both sides of the dataset, which is a pretty hard problem even without it allegedly not sharing the info on one side with Google. I guess this complaint from this article is asking the FTC to get more information on this process, or get Google to explain how it works, or something. Upvote for the explanation.

sloppycee 9y ago
You can use geography, I assume this is meant to measure ROAS. I've worked on a similar problem, and some very large retailers we worked with wanted to measure ROAS. They turned off all adspend in a certain large metropolitan area, while we turned on in that area. Measuring the sales lift/drop, gave an idea of the platform's effectiveness. Any other demographic data point could also be used, not just geo. You don't necessarily need to link people directly.
UncleMeat 9y ago
Homomorphic encryption to do set intersection. With sufficiently large sets, you don't need to do any sort of collection at the individual level.
singron 9y ago
I don't know in this case, but there are several identifiers that both parties might have (phone number, email address, billing address, name). In the industry, you can sometimes get "anonymized" data that is keyed by a hash of these identifiers, then you can hash your own identifiers and join to deanonymize. This isn't how it works in this case, but I know other companies do this.

hedora 9y ago

Assuming the encryption works as advertised, these schemes still allow more de-anonymization than you'd think. If you can compute very specific set intersections you can then get the set size low enough to reveal personal information with 100% certainty. (The set of people that bought X at M and Y at N and saw ad A at P and ad B at Q on a particular day might contain one person) I do not know of any work in homomorphic encryption that claims to provide privacy guarantees in the face of side-channels from other sources (in this case, the side-channels could be phone location information for brick-and-mortar purchases, or web browsing activity timestamps for online purchases). It's not surprising that such work is hard to find, since merely computing correct query results breaks most reasonably privacy guarantees. Also, the field of homomorphic encryption encourages people to use horrible cipher-modes like code-book, so guilt-by-association makes me highly suspicious of stuff in that space: https://en.m.wikipedia.org/wiki/Block_cipher_mode_of_operation#Electronic_Codebook_.28ECB.29 https://en.m.wikipedia.org/wiki/Block_cipher_mode_of_operati... The closest thing I know of that is workable is differential privacy, which (in laymans terms) introduces enough noise to make sure that overly-specific queries return random (but representative) garbage. Apple says they use differential privacy to train models on data that's stored on iPhones without actually giving themselves access to that data. Cool stuff.