3 ms·

Could you please clarify if the face recognition is done on-device or if the photo (or some subset of data) is uploaded to Facebook's servers in order to analys

• malandrew 11y ago
  Yes, this is hugely important. If any photos are being uploaded without my explicit action, I consider it a huge violation of trust. I've disabled every permission messenger has until there is clarification here.
  • gcr 11y ago
    The feature isn't live yet, is it?
• gcr 11y ago
  Here's one possible way they could have implemented it: All the recognition happens on the device. Faces are detected, poses are estimated and frontalized, and features are extracted on the iphone. Then, the feature vector is compared to a small gallery of your friends' features living on the telephone. No upload. That's what I hope happens. I don't really see much a problem with the above potential strategy, but then again, I wouldn't want my recognition model living on my friends' phones. (One could imagine an attack that tries to extract the images used to train the recognition model given just the model blob...are there any private images there? sounds like a juicy paper!) Or maybe the features are extracted but sent to Facebook, which returns a recognition result. I consider this less trustworthy because Facebook would learn the identity of everyone in any picture I take, even if I choose not to send the full picture to facebook. Or: perhaps faces are detected, but then cropped and the image fragment containing the face is sent to Facebook. Or more Orweillian still, perhaps the entire image is simply sent to Facebook and the recognition happens there.
  • downer70 11y ago
    You are deluded. Modern internet developers don't do this sort of thing. Not what you've described. Not as you've described it. It's never about deploying features that can't be explicitly controlled by the mothership. Furthermore, the graphics processing you're envisioning will consume battery power and might be kiboshed by the end-user's OS arbitrarily. That alone (lack of OS control) gives developers a plausible way to rationalize the orwellian. You want things to work in a way that will never ever happen.
    • gcr 11y ago
      (edit: see my reply to your other comment. I think that one's a bit better-structured to address the points you mention) I don't think it's out of the question. Here are some good reasons why Facebook might want to structure their pipeline like this. - Cellular Bandwidth is still expensive. They don't want the reputation of raising your data bill by uploading all your photos. Maybe the GPU cycles really are cheaper. After all, Facebook already uses a lot of battery anyway; what does a few million more cycles per image matter? - Messenger on iOS gets a new release every week. It's not very difficult to make code changes. Not quite as easy as a webapp, but updating recognition models is a pain anyways, so it won't happen too often. - Uploading photos without the user's consent is going to make Facebook look very bad, especially in certain countries like Germany or other parts of the EU. They already don't have a lot of social "privacy trust capital" as it is; they might not want to waste even more "trust capital" on such a silly application. Compare this to Apple, for example, who loudly claims that even they can't read your encrypted iMessages. They do this to win privacy karma points. Sure, this claim may not be true, but Facebook can't even consider this kind of marketing if photos are indeed being uploaded. On the other hand, if all the recognition happens on the device, Facebook could win some much-needed privacy karma points back by loudly saying so! Why wouldn't Facebook want to be able to say this? It's within their grasp, too! I agree; I don't think they're doing this. But my argument is much weaker: I only claim that it's possible, or perhaps even reasonable and desirable, to build their service like this.