blintz

13 ms·

121.
by blintz 4y ago
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 expen
122.
by blintz 4y ago
Not at the moment, but will probably make a blog post or something to explain how it all works at a slightly higher level than the paper. When it’s done it’ll be at https://samirmenon.com/ .
123.
by blintz 4y ago
Yeah, every record needs to be the same size. For the demo, we batch the articles into 100KB chunks. We do a good job packing, where we put many small articles into a 100KB chunk, and we split large articles into many 100KB chunks. This pac
124.
by blintz 4y ago
Yes, every item needs to be the same size. We batch the articles into 100KB chunks, and we split articles larger than this into multiple chunks.
125.
by blintz 4y ago
I do think OCSP and certificate transparency stuff is a clear application for PIR that makes a lot of sense. There, you currently basically have to volunteer your browsing patterns to a 3rd party (beyond your ISP).
126.
by blintz 4y ago
That's a good point, I could see this being a differentiating feature for a VPN provider. The only way to know if people would pay is for someone to offer it, I guess...
127.
by blintz 4y ago
The simplest approach would be to just charge people per-query (or charge in levels, depending on the number of queries). This could be done in the standard way (you have to log in, pay the site, and then it gives you an API key or just ses
128.
by blintz 4y ago
Indeed, client behavior when an article fails to be retrieved is very important. A malicious server could guess what article is being retrieved, remove it from the database, and see if the client 'reacts' in some observable way. T
129.
by blintz 4y ago
The server has a snapshot of Wikipedia, sitting in memory. The server is blind to which article is requested because it computes a homomorphic dot product between an encrypted one-hot vector (encrypted under a key that only the client kno
130.
by blintz 4y ago
The server processing has to scan the entire database to answer a query- so the entire database stays in memory, and access patterns tell you nothing about the query.
131.
by blintz 4y ago
It does sound like magic! This is what got me into this field; it seems like something that should intuitively be impossible… but it’s not! Parameters only need to be changed based on the number of items in the database (not the content of
132.
by blintz 4y ago
Sure! This paper has people really excited from a theoretical standpoint: https://eprint.iacr.org/2022/081.pdf . It’s a bit far from practical for single-server, but I’m sure people are working on it.
133.
by blintz 4y ago
There has recently been a lot of great work on making homomorphic encryption more practical for particular applications. We definitely stand on the shoulders of all that work! One reason homomorphic encryption has a reputation as absurdly s
134.
by blintz 4y ago
This is a great idea, and we think it would be relatively practical assuming some aggressive caching. However, I couldn’t think of a threat model where this is useful, since presumably your ISP can in the end always see which sites you visi
135.
by blintz 4y ago
Indeed! The encryptions of the client queries are fully semantically secure - under relatively solid lattice-based cryptographic assumptions, a server would need to do more than 2^128 work to recover the plaintext of someone’s query. One qu
136.
by blintz 4y ago
Wasn’t me, but it was accurate! Indeed, except in some (exciting!) new theoretical constructions, server work is always linear in the database size. However, I’d emphasize that our work shows that the constabt factor on this linear operatio
137.
by blintz 4y 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 com
138.
Show HN: Read Wikipedia privately using homomorphic encryption(spiralwiki.com)
331 points by blintz 4y ago|116 comments
139.
by blintz 4y ago
These things are indeed possible while maintaining fully semantically secure encryption. Recent, mostly theoretical work shows that this is possible using fully homomorphic encryption. The basic idea is, the client can encrypt its query, th
140.
by blintz 4y ago
The detail that they also reassured the people with offers that their offers would not be rescinded two weeks ago is wild. Did they expect this not to leak? Why weren’t they transparent about this total about face in the initial announcem
141.
by blintz 4y ago
Discussion forums like Piazza can really help with stuff like this. By allowing students to communicate with each other in public, with TAs/instructors explicitly present answering questions, the need to get a group chat going diminish
142.
by blintz 4y ago
This is maybe a bit of a tangent, but the fact that transaction validation happens as part of Bitcoin’s proof-of-work process is incidental. It’s perfectly fine to mine an empty block, and the only incentive to include transactions is trans
143.
by blintz 4y ago
You're right in some sense, I mean, software engineering is a skill that lots of companies outside of the video game industry will pay quite well for. Still, it seems like there is a large oversupply of people who really want to work s
144.
by blintz 4y ago
This is why I think the parallel to the movie industry is so useful: there, you also have actors/production staff/etc willing to work for essentially nothing, and an industry that, without the unions, would be more than happy to e
145.
by blintz 4y ago
I suppose this is run-of-the-mill for the video game industry, but man, it's sounds so hellish; people deserve not be humiliated like this at work. Isn't the solution unions, akin to the heavily unionized movie / TV business?
146.
by blintz 5y ago
I have often wondered how close the fees that 'marketplace' companies charge are to the value they provide; of course, in a perfectly competitive world, they would be very close. However, it seems that in practice, the marketplace
147.
by blintz 5y ago
Very cool that this is just from logic bugs! I wonder if we should as a rule assume that sandboxing that is not formally verified or battle tested over a really long time is unlikely to be free of bugs. What's the long-term solution fo
148.
by blintz 5y ago
Is there any plan to just publish your list of keywords? While you don’t collect any PII, to use Koko, it seems like you still have to send all content you want to scan to your API. For things like DMs or private posts, this seems less than
149.
by blintz 5y ago
San Francisco has a lower violent crime rate than Houston, Dallas, Washington D.C., and Miami. https://en.wikipedia.org/wiki/List_of_United_States_cities_b... Is it possible that these other cities also have violence,
150.
by blintz 5y ago
Yup: "Now we're going to have designers from different backgrounds, different voices, different experiences, being able to have a seat at the table." (34:30)

More ›