Hacker News Reader

Comment by j2kun | Hacker News Reader

There is a possibility to share statistical analysis of data privately. This is the subfield of computer science known as differential privacy, and I bet companies just looking for trend data (not specific targeting) would be able to benefit from user data without compromising user privacy.

In other words, there are good and bad ways to use user data, some of the good ones are provably private, and we shouldn't rule out all data sharing as inherently evil.

View on HN

That's all fine, but there is zero evidence that any company is doing that.

Furthermore, even if a company was using differential privacy to anonymize data between the collection and the processing step, there is no way to prove that the data is being handled appropriately before and during the collection step. Also, there is no way for a company to prove that they continue to handle data appropriately over time.

These tools are super useful for releasing data sets to the public from a trusted source (such as a healthcare provider assisting researchers by releasing cancer data), but that's not what we're talking about here.

It is a new field, I admit, but it shows there's a false dichotomy between privacy and stopping all data collection and analysis.

LeapYear Innovations [1] looks like they have some clients who care about differential privacy. [2]

[1]: http://www.leapyearinnovations.com/

[2]: http://www.leapyearinnovations.com/shroudbase

It's not a false dichotomy: most of the time data collection and analysis does not preserve privacy. In all cases, you can't guarantee that.

You also didn't respond to this:

> Furthermore, even if a company was using differential privacy to anonymize data between the collection and the processing step, there is no way to prove that the data is being handled appropriately before and during the collection step. Also, there is no way for a company to prove that they continue to handle data appropriately over time.

Any company using shroudbase can choose to stop using shroudbase at any time, or accidentally introduce a bug into their collection method, or start also saving the data in MySQL at collection time.

Tools like that are useful to distribute data from trusted to non-trusted parties. You still have to trust the analytics and advertising companies doing the collection.

You also can't prove that someone won't poison your soup before, during, or after it was prepared from raw ingredients. But you don't claim either to avoid all soup or die. This is because you can give very good guarantees that if the correct process is followed then that won't happen. This is pretty much how we get by in all areas of life. Trust is a default and it has to be for society to function. If we can back up that trust by provably good protocols, then one must be malicious in order to break the system.

Your chef doesn't have a large monetary incentive to poison your soup. Would you eat the food if he did?

They have large monetary incentives to cut corners when it comes to QA and picking a distributor. One big thing stopping them is oversight, which is perfectly reasonable to impose on data collection policies. But what do you need to do so? Proof that the imposed policies give a good guarantee.

>we shouldn't rule out all data sharing as inherently evil.

I agree. However, that's going to be the knee-jerk reaction for a while precisely because huge entities like Google, NSA, AT&T, Verizon, etc have been doing that data sharing while doing an abysmal job of notifying users about what was happening. Eventually I think we'll settle into a happy medium, but we're in the pushback phase at the moment.

> provably private

citation?

http://research.microsoft.com/pubs/74339/dwork_tamc.pdf

They define the mathematical guarantees on privacy, which are computational guarantees against a polynomial-time adversary trying to distinguish between the analyses of two databases that differ in a single record. I.e., nobody can reasonably tell whether your information is included in the analysis.

I am not particularly looking for a debate today; but this seems to be a disagreement of terms when using the word "provable". Just because a polynomial time algorithm can not be constructed to distinguish a single record does not mean private information can not be derived.

I still would prefer a choice in the matter - a required small check box that says "record my data" or "do not record my data" XOR "erase my data". The term 'reasonably' has a very different meaning to me (assuming many unknowns in the future) over the long term.

It is either that, or people should literally stop making such a hard line distinguishing person A from person B. If the data is going to be used to make an inference from the collective to the individual, and it doesn't matter whether my data is included or not, well, shrug and thumbs up. It really depends on the context of the application. Does it determine whether my imaginary future children get to go to college or not?

The problem with data and private organizations is that we do not know what the data analysis is being used for, and as private citizens we do not have any control over how that data is mathematically reasoned about and qualitatively assessed aside from forging the data itself it (garbage in, garbage out).

They define what they mean by the word private, and then prove their mechanisms satisfy that definition. It's not a disagreement about the word "provable," it's that you don't agree that their definition of privacy is a good definition of privacy. That's fine, but your reason seems suspicious. If they could determine any information about you, then such an algorithm would fail if your data were not collected in the first place. Being able to derive private information implies the ability to distinguish.

And I misspoke about polynomial time adversaries. Reading closer it looks like an information theoretic guarantee.

> That's fine, but your reason seems suspicious.

I prefer to avoid living my life in a fishbowl.

> provably private

This does not have a mathematics definition. Private means not accessible to anyone aside from the owner. That is something that is provably private.

> information theoretic guarantee

This sounds like hand waving to get rid of actual concerns.

Again, you are disagreeing with their chosen definition for a word you prefer to think of in a different way. Maybe you should study the issue instead of griping about how lame their version is. What does accessible mean? Who is the owner of a piece of data? For example, just because of our interaction here, I could probably get a pretty good estimate of your occupation, age, and gender. Is that "accessing" something? If so, does that mean speaking is enough to break privacy? Do you "own" your age and gender? If this conversation is enough for me to determine your age and gender, do you own this conversation?

Definitions of words matter when people skim text as to determine how they communicate their opinions, form their opinions, and take action, including writing legislation.

If I was actually typing out the opinions of my spouse word for word: whose age combined with my own is divisible by 3, how would that change your view?

Of course definitions matter. My point is that your definition is not specific enough to be useful.

Yours is mathematically nonsensical. You either have discrete containers that denote privacy relations between information and parties or you do not have it at all. Privacy does not exist on a gradient on a computer. You can not calculate the percentage of how private something is. If it is marked 'private information' and that information is released to a party not under the user / group / role term that is supposed to have access to that category of information, then that information ceases to be private.

You're discussing a different problem. You're using "private" in a legal sense. As in "it's either legal or it's not". Passing privacy laws is certainly a good idea and is what the DDG post is advocating. j2kun was bringing up another solution to privacy which was cryptographic rather than legal. Cryptographers generally use "private" in a "is it physically possible for a bad actor to decrypt this" sense. In which case, there's always a possibility that bad actor can access it. It's just a question of "how likely?" (any encryption requiring a deciphering-technique which takes more than polynomial-time to crack is considered good-enough).

In order for the data to be of any use, someone (be it the user or a trusted friend) has to be able to decrypt it. But if you (or your friend) can decrypt it, a hacker can also decrypt it. Hopefully, the hacker will find it more difficult without knowing the key. But the one in a million chance that he'll crack it on the first try exists by necessity.Even titanium breaks under stress.

In the pathological case that there's zero possibility (not even epsilon), the whole business of encryption is rendered moot because the rightful user himself cannot access the data! This means one of two things: the data is corrupt; or the data has been deleted entirely.

Computer security has more subfields than cryptography.

http://en.m.wikipedia.org/wiki/Information_flow_%28informati...

Again, you are revealing your ignorance of mathematics and arguing that you don't like the way a word is used in a context you don't understand.

Okay, but those ad hominem attacks won't stop some people from training their machines better than implementations of these algorithms can manage the complexities between data kinds and data object noise.

If you think privacy is "power over what others perceive", then you are being selfish and short-sighted in a way that prevents mutual privacy, and is thus generally untenable.

That would be fine, if the 'people who perceive' would realize perception is much more variant and undecidable than 'people who perceive' acknowledge.

Reply on news.ycombinator.com