3 ms·

"Ok Google" on Android, Siri, and listen all the time too, don't they? That's presumably how I can wake up my phone from a state where the screen is locked and

by Pyxl101 11y ago

"Ok Google" on Android, Siri, and listen all the time too, don't they? That's presumably how I can wake up my phone from a state where the screen is locked and powered off by saying "Hey Siri" / "Ok Google" / "Hey Cortana". I was taken aback briefly when I first considered it. I realized it must be recording the environment constantly, in order to scan for the code word. Upon hearing the code word, it wakes up and streams the last few seconds of audio remotely, where the real linguistic analysis happens with Siri/Cortana/Google Voice Search. [1] The difference is my phone is with me virtually 24/7. It's rarely not in the same room. In my opinion, the cat was out of the bag a while ago. I make a decision to trust the device manufacturers, who claim their devices only listen for the code word. As a scientist and engineer, I know what that means: it's recording everything, all the time, and probably running a simple linguistic analysis sufficient to recognize codewords only. Then on waking up, it streams that audio centrally, which is how it can respond to queries out of the blue like: "Hey Siri, what's the temperature?" Answer: "It's currently 41 degrees Fahrenheit". It must be recording and processing everything, though in a computationally inexpensive way, for the sake of battery life. [1] Google Chrome apparently supported this on the desktop for a while too: http://thenextweb.com/google/2013/11/26/google-brings-ok-google-hotword-voice-search-desktop-new-chrome-extension/ http://thenextweb.com/google/2013/11/26/google-brings-ok-goo...

acgourley 11y ago
In pursuit of a low battery foot print and low memory foot print the "wake up word" technology is really just a trained ML classifier running locally to identify a specific 3+ syllable phrase. They don't start sending everything to the mothership until after that gate is passed.
- x5n1 11y ago
  What happens when they do?
  - xgbi 11y ago
    Your phone battery dies in 3 hours. This is why we're "lucky" to have limited battery tech right now. What's to fear is when a minuscule device will get years and years of energy with new battery tech. This is where ubiquitous connected devices will ramp up and pose the real privacy threat.
    - lolive 11y ago
      I will happily share all my secrets with such devices. Just give them the voice of Scarlett Johansson :)
    - nkozyra 11y ago
      > Your phone battery dies in 3 hours. I'm not so sure. Passive background network activity happens perpetually on Android and iOS devices now. I'm not saying it doesn't have an impact on battery life, I'm just saying that sending chunks of speech during this time wouldn't have significantly more impact than Gmail syncing in the background.
    - biot 11y ago
      My iPhone has Hey Siri on all the time and lasts for 2-3 days between charges.
      - xgbi 11y ago
        Because it does all voice recognition in-place by a highly specific chip. If you send out all voice data to a server, you're tapping into both CPU and radio, which is quite equivalent as when you speak on, say, Facetime audio or a VoIP app. I dare you to speak for more than a few hours with WhatsApp.
- RockyMcNuts 11y ago
  So when it's plugged in to power it can send everything? how reassuring, since I pretty much always plug it in when I'm home? Related, my 'Hey Siri' is only active when the iPhone is plugged in.
  - scep12 11y ago
    It could send everything, but you can do some pretty basic investigation to figure out if it actually does. Spoiler: It doesn't.
dasil003 11y ago
It's probably naive, but somehow I feel better about the phone because it has power and bandwidth limitations that would quickly become apparent if it started trying to send all this data somewhere. I'm much more concerned about devices that are wired in permanently with power and network.
- MichaelApproved 11y ago
  As technology advances, this will not be as comforting as it is now. Also, the energy issue is irrelevant while the phone is plugged in to charge. Will it change your phone use? edit typo
- Pyxl101 11y ago
  That's true. Wired devices have the power to listen passively indefinitely and to record as much as they want. As I updated my post to mention, Google apparently added Google Voice search to Chrome at one point. With wired devices, I've comforted myself with the analysis that users could detect subterfuge by looking for unexpected network traffic. A device can't record my voice and transmit it centrally all the time without generating noticeable network usage. It could, however, record for a while and send in a burst later, perhaps concealed within other traffic like a weekly software update. Nevertheless, I think a close examination could detect foul play. I do wonder whether phones could conduct the same type of subterfuge: Perhaps there ought to be a "phone surveillance challenge" that's kind of like the Underhanded C contest. Each year, contestants submit a phone app that runs in the background on a phone over a 48 hour contest period simulating normal usage. At several points during that period, an important conversation will be overheard. The app that does the best job recording the conversations and remaining undetected is the winner. Naively recording all audio and storing it will probably drain the battery. It also needs to be compressed, though bit rates as low as 5 Kbps seem to be intelligible. If we record at 5 Kbps, then it will require only 108 megabytes to store 48 hours worth of audio. The compression might eat up CPU, however. Perhaps skilled contestants will find a way to do it cheaply. If we're a smart contestant, we don't indiscriminately record and store all audio. Instead, we want our contest entry to record the human conversations. I'd hazard a guess that our highly skilled contestants can come up with cheap passive methods of detecting human voice, comparable to the ones that listen for "Ok Google". So our contest entry only wakes up when it hears human speech. Our contest entry won't transmit anything over the airwaves regularly, since that uses battery. However, we know the phone will be charged at least once in 48 hours. We'll wait until it's charging to upload our recordings. Even better, we'll time our transmission burst to be right after some other network activity, like a daily software update. Or if the phone is plugged into a computer, we can compromise it as well and extract data that way. Sounds like a fun contest.

by Pyxl101 11y ago

acgourley 11y ago

In pursuit of a low battery foot print and low memory foot print the "wake up word" technology is really just a trained ML classifier running locally to identify a specific 3+ syllable phrase. They don't start sending everything to the mothership until after that gate is passed.

x5n1 11y ago
What happens when they do?
- xgbi 11y ago
  Your phone battery dies in 3 hours. This is why we're "lucky" to have limited battery tech right now. What's to fear is when a minuscule device will get years and years of energy with new battery tech. This is where ubiquitous connected devices will ramp up and pose the real privacy threat.
  - lolive 11y ago
    I will happily share all my secrets with such devices. Just give them the voice of Scarlett Johansson :)
  - nkozyra 11y ago
    > Your phone battery dies in 3 hours. I'm not so sure. Passive background network activity happens perpetually on Android and iOS devices now. I'm not saying it doesn't have an impact on battery life, I'm just saying that sending chunks of speech during this time wouldn't have significantly more impact than Gmail syncing in the background.
  - biot 11y ago
    My iPhone has Hey Siri on all the time and lasts for 2-3 days between charges.
    - xgbi 11y ago
      Because it does all voice recognition in-place by a highly specific chip. If you send out all voice data to a server, you're tapping into both CPU and radio, which is quite equivalent as when you speak on, say, Facetime audio or a VoIP app. I dare you to speak for more than a few hours with WhatsApp.
RockyMcNuts 11y ago
So when it's plugged in to power it can send everything? how reassuring, since I pretty much always plug it in when I'm home? Related, my 'Hey Siri' is only active when the iPhone is plugged in.
- scep12 11y ago
  It could send everything, but you can do some pretty basic investigation to figure out if it actually does. Spoiler: It doesn't.

dasil003 11y ago

It's probably naive, but somehow I feel better about the phone because it has power and bandwidth limitations that would quickly become apparent if it started trying to send all this data somewhere. I'm much more concerned about devices that are wired in permanently with power and network.

MichaelApproved 11y ago
As technology advances, this will not be as comforting as it is now. Also, the energy issue is irrelevant while the phone is plugged in to charge. Will it change your phone use? edit typo
Pyxl101 11y ago
That's true. Wired devices have the power to listen passively indefinitely and to record as much as they want. As I updated my post to mention, Google apparently added Google Voice search to Chrome at one point. With wired devices, I've comforted myself with the analysis that users could detect subterfuge by looking for unexpected network traffic. A device can't record my voice and transmit it centrally all the time without generating noticeable network usage. It could, however, record for a while and send in a burst later, perhaps concealed within other traffic like a weekly software update. Nevertheless, I think a close examination could detect foul play. I do wonder whether phones could conduct the same type of subterfuge: Perhaps there ought to be a "phone surveillance challenge" that's kind of like the Underhanded C contest. Each year, contestants submit a phone app that runs in the background on a phone over a 48 hour contest period simulating normal usage. At several points during that period, an important conversation will be overheard. The app that does the best job recording the conversations and remaining undetected is the winner. Naively recording all audio and storing it will probably drain the battery. It also needs to be compressed, though bit rates as low as 5 Kbps seem to be intelligible. If we record at 5 Kbps, then it will require only 108 megabytes to store 48 hours worth of audio. The compression might eat up CPU, however. Perhaps skilled contestants will find a way to do it cheaply. If we're a smart contestant, we don't indiscriminately record and store all audio. Instead, we want our contest entry to record the human conversations. I'd hazard a guess that our highly skilled contestants can come up with cheap passive methods of detecting human voice, comparable to the ones that listen for "Ok Google". So our contest entry only wakes up when it hears human speech. Our contest entry won't transmit anything over the airwaves regularly, since that uses battery. However, we know the phone will be charged at least once in 48 hours. We'll wait until it's charging to upload our recordings. Even better, we'll time our transmission burst to be right after some other network activity, like a daily software update. Or if the phone is plugged into a computer, we can compromise it as well and extract data that way. Sounds like a fun contest.