48 

    49 %Differential privacy

    50 %=====================

    51 %https://www.wired.com/2016/06/apples-differential-privacy-collecting-data/

    52 

    53 %Differential privacy, translated from Apple-speak, is the

    54 %statistical science of trying to learn as much as possible

    55 %about a group while learning as little as possible about any

    56 %individual in it.

    57 

    58 %As Roth notes when he refers to a “mathematical proof,”

    59 %differential privacy doesn’t merely try to obfuscate or

    60 %“anonymize” users’ data. That anonymization approach, he

    61 %argues, tends to fail. In 2007, for instance, Netflix released

    62 %a large collection of its viewers’ film ratings as part of a

    63 %competition to optimize its recommendations, removing people’s

    64 %names and other identifying details and publishing only their

    65 %Netflix ratings. But researchers soon cross-referenced the

    66 %Netflix data with public review data on IMDB to match up

    67 %similar patterns of recommendations between the sites and add

    68 %names back into Netflix’s supposedly anonymous database.

    69 

    70 %As an example of that last method, Microsoft’s Dwork points to

    71 %the technique in which a survey asks if the respondent has

    72 %ever, say, broken a law. But first, the survey asks them to

    73 %flip a coin. If the result is tails, they should answer

    74 %honestly. If the result is heads, they’re instructed to flip

    75 %the coin again and then answer “yes” for heads or “no” for

    76 %tails. The resulting random noise can be subtracted from the

    77 %results with a bit of algebra, and every respondent is

    78 %protected from punishment if they admitted to lawbreaking.

    79 

    80 %https://www.cis.upenn.edu/~aaroth/Papers/privacybook.pdf