4 ms·

That's kind of like saying "how would you make a car without wheels?". It could be that inherent in the idea of making object X are privacy concerns. The fact t

That's kind of like saying "how would you make a car without wheels?". It could be that inherent in the idea of making object X are privacy concerns. The fact that the privacy concerns are intrinsic to the object just means that you have to call the whole object X into question when discussing them. Your question seems to suggest a sentiment like "there isn't any other way to do this, can you think of one?" but there is, and it's not to do it at all. (I'm not taking a position on the Nest device, but this comment just seemed a little like a cognitive bias similar to anchoring)

dguaraglia 11y ago
Exactly! And I'm OK with people taking the stance "if it's not 100% private and my video needs to be sent over the internet, then I don't want the camera". You can either have a camera that keeps a safe copy of your video in the cloud and detects unexpected activity happening in front of your camera, or you can have one that is 100% private and doesn't upload video to the cloud. You can't have both. At least not without having a huge setup in a secured room inside your home. I understand the privacy concerns, I just argue that you can't have the ideal service without getting over them. At least not with current technology, and probably not without the right economic incentives to build a stand-alone system.
- bloomingfractal 11y ago
  What about doing all processing locally and storing only end-to-end encrypted data in servers?
  - dguaraglia 11y ago
    Not really an option right now. The current generation of cameras out there are basically a Raspberry-Pi level computer with a better camera and a hardware h264 encoder. The moment you start doing something fancy, like running any non-trivial motion-detection algorithm, you are bound to run into performance or thermal (read: overheating) issues. Let's not even talk about machine learning. Just think how much money Nest would save in server time with such a setup :)
    - bloomingfractal 11y ago
      I see, Thanks for the insightful answer! Regarding machine learning it would be nice to be able do the training in your desktop/laptop when it's idle or something like that. But it's very good to know this is a technological issue (as opposed to a business issue). Well, hope you smart folks solve this. Meanwhile, I'll keep tinkering with my raspberry pi and raspberry pi camera :)
    - lazaroclapp 11y ago
      My intuition says we are not anywhere near there yet, but do you know if any video processing algorithms exist that can reasonably be executed on encrypted data? Basically, I know fully-homomorphic encryption is ridiculously inefficient in the general case. At the same time, I know of specialized homomorphic encryption algorithms that can operate on encrypted data of specific formats. There are efficient-ish algorithms for encrypted (social-network-type) graphs, and encrypted vote ballots. I was wondering if you or anyone in your team has come across any work on privacy-preserving encrypted audio/image/video processing? I assume this is a very hard problem, but I imagine someone has tried looking into it.
      - dguaraglia 11y ago
        I'm not really the person to answer that (I'm a lowly software engineer keeping the cogs greased m'lord!) I know at some point Larry (https://github.com/lwneal https://github.com/lwneal) was looking into that, at least cursorily. I'll refer to him as the authority on anything encryption-related at Dropcam (or anything, in general. Brilliant guy!)
    - detaro 11y ago
      Can you share any reference on what the Nest servers actually do that a smartphone chipset isn't capable of? Various apps manage (from my limited knowledge about the field) quite impressive things.
      - dguaraglia 11y ago
        Can't really discuss specifics, but training machine learning models that share data between all your cameras would be pretty difficult, for example. Also, the lifespan of the processor would be highly reduced if you were constantly hammering it (thermal implications, etc. etc.) Again, in a few years that might not be an issue anymore :)

dguaraglia 11y ago

Exactly! And I'm OK with people taking the stance "if it's not 100% private and my video needs to be sent over the internet, then I don't want the camera". You can either have a camera that keeps a safe copy of your video in the cloud and detects unexpected activity happening in front of your camera, or you can have one that is 100% private and doesn't upload video to the cloud. You can't have both. At least not without having a huge setup in a secured room inside your home. I understand the privacy concerns, I just argue that you can't have the ideal service without getting over them. At least not with current technology, and probably not without the right economic incentives to build a stand-alone system.

bloomingfractal 11y ago
What about doing all processing locally and storing only end-to-end encrypted data in servers?
- dguaraglia 11y ago
  Not really an option right now. The current generation of cameras out there are basically a Raspberry-Pi level computer with a better camera and a hardware h264 encoder. The moment you start doing something fancy, like running any non-trivial motion-detection algorithm, you are bound to run into performance or thermal (read: overheating) issues. Let's not even talk about machine learning. Just think how much money Nest would save in server time with such a setup :)
  - bloomingfractal 11y ago
    I see, Thanks for the insightful answer! Regarding machine learning it would be nice to be able do the training in your desktop/laptop when it's idle or something like that. But it's very good to know this is a technological issue (as opposed to a business issue). Well, hope you smart folks solve this. Meanwhile, I'll keep tinkering with my raspberry pi and raspberry pi camera :)
  - lazaroclapp 11y ago
    My intuition says we are not anywhere near there yet, but do you know if any video processing algorithms exist that can reasonably be executed on encrypted data? Basically, I know fully-homomorphic encryption is ridiculously inefficient in the general case. At the same time, I know of specialized homomorphic encryption algorithms that can operate on encrypted data of specific formats. There are efficient-ish algorithms for encrypted (social-network-type) graphs, and encrypted vote ballots. I was wondering if you or anyone in your team has come across any work on privacy-preserving encrypted audio/image/video processing? I assume this is a very hard problem, but I imagine someone has tried looking into it.
    - dguaraglia 11y ago
      I'm not really the person to answer that (I'm a lowly software engineer keeping the cogs greased m'lord!) I know at some point Larry (https://github.com/lwneal https://github.com/lwneal) was looking into that, at least cursorily. I'll refer to him as the authority on anything encryption-related at Dropcam (or anything, in general. Brilliant guy!)
  - detaro 11y ago
    Can you share any reference on what the Nest servers actually do that a smartphone chipset isn't capable of? Various apps manage (from my limited knowledge about the field) quite impressive things.
    - dguaraglia 11y ago
      Can't really discuss specifics, but training machine learning models that share data between all your cameras would be pretty difficult, for example. Also, the lifespan of the processor would be highly reduced if you were constantly hammering it (thermal implications, etc. etc.) Again, in a few years that might not be an issue anymore :)