5 ms·

I do data mining. Ready here we go. 10% of people who used service discovered they have BrandonWirtz's Disease. Requirements: Male. Primarily Caucasion as pr

by drakaal 14y ago

I do data mining. Ready here we go. 10% of people who used service discovered they have BrandonWirtz's Disease. Requirements: Male. Primarily Caucasion as primarily limited to Arian Descent. Symptoms onset at 55. Not symptomatic for those with BMI under 22. With this information I can take the 100% and cut the list to 50% that are male. 50% that are caucasion. (25%) Would have already know if they were over 55 so we cut another 25% Strip last names for ones that would not be of the right heritage. Suddenly the eligible candidates are only 12% of the total. I know that 80% of those are the people on that list. That's how mining works. You can't share data anonymously becuase each facet of data allows me more information to piece things together. Sharing has to be opt in. But if it is opt in then it is no longer statiscitally relevant. Because just as an example. People are going to be less apt to share data if they have a disease associated with a stigma. Which will skew the results. [Brandon Wirtz's Disease is an imaginary disease used for demonstration only. I don't have any genes shared with other humans as I am from another planet.]

lightcatcher 14y ago
I'm currently taking a class in data privacy, which is about assessing privacy risks using a mathematically rigorous formalization (differential privacy, see http://en.wikipedia.org/wiki/Differential_privacy http://en.wikipedia.org/wiki/Differential_privacy ) and developing algorithms to allow for minimal risk release of private databases and avoid attacks like yours above. Your post makes it seem like you might enjoy reading more about the field; you should check out this survey of the field: http://www.cs.ucdavis.edu/~franklin/ecs289/2010/dwork_2008.pdf http://www.cs.ucdavis.edu/~franklin/ecs289/2010/dwork_2008.p...
- drakaal 14y ago
  Those formula's assume the creator knows all the variables. I have data on things like Probablity a First name is of which race. When names started to appear for determining maximum Age. Search data for identifying people who were worried they had a disease. Often a single piece of data will cause the intersection of two traits to be 80% accurate. You can't mathmatically predict these things because you don't know how many factors I have. There are zip codes where only 1 in 100 people are of a given race. There are diseases that only 1 in 100k people have done a search for and only 1 in 500k people have. These kinds of things you can't formulaize.
  - lightcatcher 14y ago
    This is not true at all. Differential privacy works to counteract EXACTLY these types of attacks. There are some very basic proofs that the privacy guarantees of a differential private release mechanism is not at all weakened by external data, such as what you've listed. Differential privacy attempts to minimize the risk of being a row in a database given that the attacker knows the values of every other row (and has whatever external data he wants).
pyre 14y ago
Maybe I'm missing something. How does "10% of our customers have X disease" allow you to get down to a specific person? | With this information I can take the 100% and cut the list | to 50% that are male. 50% that are caucasion. (25%) Would | have already know if they were over 55 so we cut another | 25% Strip last names for ones that would not be of the right | heritage. Suddenly the eligible candidates are only 12% of | the total. | I know that 80% of those are the people on that | list. What list? Are you somehow taking a list of names from outside and correlating them based just on 'X%'?
- drakaal 14y ago
  Things like cookie inspection allow a data miner to get a list of known customers of a company. I can tell you every RackSpace Customer, Every Amazon Shopper, Every FaceBook User. If you have seen those ads on FB that are for something you just searched on Amazon, or that want to sell you hosting for a place you already have hosting with you know what I mean. If you have enough data about the factors that are common of people with a disease, and then compare that with percentage of people known to have that disease from a set. If 10% of Company's American clients have Disease X, and Disease X requires you live in a cold climate to be diagnosed ( Raynaud's for example ) then you can exclude the people who live in a warm climate. Now if you want to target people with Raynauds and you know 80% of the clients live in California,Texas, And Nevada. You know that 50% of the remaining clients have Raynaud's. This is of course a contrived example, and uses much less data than would be used by a Data Miner. For the people saying "no there are formula's we can prevent this from happening" yeah you know who made those formula's? The People who make money selling data. And thos formulas are to keep you from thinking it is dangerous.

by drakaal 14y ago

lightcatcher 14y ago

I'm currently taking a class in data privacy, which is about assessing privacy risks using a mathematically rigorous formalization (differential privacy, see http://en.wikipedia.org/wiki/Differential_privacy http://en.wikipedia.org/wiki/Differential_privacy ) and developing algorithms to allow for minimal risk release of private databases and avoid attacks like yours above. Your post makes it seem like you might enjoy reading more about the field; you should check out this survey of the field: http://www.cs.ucdavis.edu/~franklin/ecs289/2010/dwork_2008.pdf http://www.cs.ucdavis.edu/~franklin/ecs289/2010/dwork_2008.p...

drakaal 14y ago
Those formula's assume the creator knows all the variables. I have data on things like Probablity a First name is of which race. When names started to appear for determining maximum Age. Search data for identifying people who were worried they had a disease. Often a single piece of data will cause the intersection of two traits to be 80% accurate. You can't mathmatically predict these things because you don't know how many factors I have. There are zip codes where only 1 in 100 people are of a given race. There are diseases that only 1 in 100k people have done a search for and only 1 in 500k people have. These kinds of things you can't formulaize.
- lightcatcher 14y ago
  This is not true at all. Differential privacy works to counteract EXACTLY these types of attacks. There are some very basic proofs that the privacy guarantees of a differential private release mechanism is not at all weakened by external data, such as what you've listed. Differential privacy attempts to minimize the risk of being a row in a database given that the attacker knows the values of every other row (and has whatever external data he wants).

pyre 14y ago

Maybe I'm missing something. How does "10% of our customers have X disease" allow you to get down to a specific person? | With this information I can take the 100% and cut the list | to 50% that are male. 50% that are caucasion. (25%) Would | have already know if they were over 55 so we cut another | 25% Strip last names for ones that would not be of the right | heritage. Suddenly the eligible candidates are only 12% of | the total. | I know that 80% of those are the people on that | list. What list? Are you somehow taking a list of names from outside and correlating them based just on 'X%'?

drakaal 14y ago
Things like cookie inspection allow a data miner to get a list of known customers of a company. I can tell you every RackSpace Customer, Every Amazon Shopper, Every FaceBook User. If you have seen those ads on FB that are for something you just searched on Amazon, or that want to sell you hosting for a place you already have hosting with you know what I mean. If you have enough data about the factors that are common of people with a disease, and then compare that with percentage of people known to have that disease from a set. If 10% of Company's American clients have Disease X, and Disease X requires you live in a cold climate to be diagnosed ( Raynaud's for example ) then you can exclude the people who live in a warm climate. Now if you want to target people with Raynauds and you know 80% of the clients live in California,Texas, And Nevada. You know that 50% of the remaining clients have Raynaud's. This is of course a contrived example, and uses much less data than would be used by a Data Miner. For the people saying "no there are formula's we can prevent this from happening" yeah you know who made those formula's? The People who make money selling data. And thos formulas are to keep you from thinking it is dangerous.