Comment by blintz
3 years ago
The $35/month server uses 6 cores to answer a single query in ~2.5-3 seconds. So it’s 0.33 QPS :-)
Not high, which is why it might not be working well for folks right now…
Time scales almost perfectly linearly with cores, since the computation is embarrassingly parallel.
In terms of cost, we’re still talking only 18 CPU•s and ~300KB of outgoing bandwidth, which is not a ton at todays prices.
It's embarrassing parallel... but you also do N times more work than a non-homomorphic system so that's not saying much!
This doesn't seem like a particularly compelling application - can you give some practical problems that homomorphic encryption solves. I've always heard vote counting as the example.
Yeah, the computational overhead over no-privacy retrieval will always be high. Something we point out in the paper is that this does not necessarily translate to server monetary costs; for example, on AWS, where outgoing bandwidth is expensive, using our system to stream a movie is less than 2x the monetary cost of direct, no-privacy streaming. This is because the computational costs are just very small in comparison to the bandwidth costs.
As far as more compelling applications, several folks have suggested DNS, OCSP, and certificate transparency as potential examples. Using PIR as a building block, it is also possible to build more complex systems, like metadata-private messaging (systems like ‘Pung’) and even private voice calling (systems like ‘Addra’).