3 ms·
In past discussions, I was led to believe we were safe from these devices, in terms of pervasive data collection, because the main processor that was always run
by z_open 9y ago
In past discussions, I was led to believe we were safe from these devices, in terms of pervasive data collection, because the main processor that was always running was primitive and was only powerful enough to pick up a key phrase. This phrase would power the main processor which would then listen to your query, respond, and then deactivate. But now, as the article mentions, I realize that this initial processor could be made to listen for many key words like "love", "hate", or other words that would pick up on sensitive personal information. I really don't think these devices should exist.
- pdkl95 9y ago> because the main processor that was always running was primitive and was only powerful enough to pick up a key phrase That's only true if you're trying to do complex analysis on the device. Cell phones in the early 1990s had even less computing power, but it was enough to encode speech down to 13.2 kbit/s[1] (or 5.6 kbit/s[2]). A simple noise gate[3] would reduce recording duty cycle down to maybe 1% while costing a trivial amount of CPU load. Modern hardware - even tiny embedded devices - can probably do a lot more than simple gat4ed+compressed audio. [1] https://en.wikipedia.org/wiki/Full_Rate https://en.wikipedia.org/wiki/Full_Rate [2] https://en.wikipedia.org/wiki/Half_Rate https://en.wikipedia.org/wiki/Half_Rate [3] https://en.wikipedia.org/wiki/Noise_gate https://en.wikipedia.org/wiki/Noise_gate
- qvrjuec 9y agoSo let's assume this is technically possible, and only examine the process under the lens of detectability. Assuming the codec is 8kbps, you'd see roughly 3MB upload for every hour of audio, and I would say it's not out of the question for the noise gate to be active from the television being on in the same room, or music playing. For an utterance, seeing a 3MB upstream would be super abnormal. It would be impossible to transmit all of this data for processing without somebody noticing.
- jononor 9y agoWhy would it record continiously? If I were developing such a device I'd use random sampling. Don't have use for all the audio in every home, but a little bit from everywhere can be useful. Can be used to improve the acoustic model of the room maybe, or improve targetting of ads[1]. Some snippets from the minutes before and activation could be especially interesting, to understand the context. 1. Or more nefarious things, of course.
- soared 9y agoOr record audio, do speech to text, some light analysis, and upload the metadata. Or just upload a basket of words and their frequency, "Toilet Paper, 2, tv, 7, online, 4,"
- jononor 9y agoDetecting who speaks is also a (relatively) lightweight analysis. Combined with bag-of-words can build personality/interest profile. Speech based gender detection can also be done, probably also detecting kids from adults. Now have good data about demographics of the household.
- pdkl95 9y agoWhile a noise gate is an obvious starting point, I suspect that much smarter filters are possible. A wide range of design choices are possible that trade CPU <-> complexity <-> accuracy. If a high error rate (including "no data") is acceptable, a filter might simply default to "off" when any unusual condition like "noise gate has been on continuously for 15 minutes" is detected. If I was designing this kind of spyware, I would put a hard limit that cuts off uploading any "extra" data after sending some low multiple of the "legitimate" data.