CSC6585 2024 Privacy 3

Поділитися
Вставка
  • Опубліковано 21 лис 2024

КОМЕНТАРІ • 2

  • @LycanB64
    @LycanB64 15 днів тому

    When reverse engineering the spinner for differential privacy and trying to get a privacy budget of 1.1, why devote 50% of the spinner to yes and cut down the 'no', instead of cutting the same proportion out of each?

    • @StacyProwell
      @StacyProwell  14 днів тому

      Great question! On the slide Reverse Engineering the Spinner (1) we need two probabilities: the probability of observing YES when the true state is YES, and the probability of observing YES when the true state is NO.
      I think what you are asking is whether the proportion that goes to NO is arbitrary. It is. I just have to divide the space between YES and TRUTH so that the correct ratio is achieved.
      Let's say I have three areas: AT, AY, and AN for TRUTH, YES, and NO, respectively. Like on the slide I will work out that Pr[YES|YES] = AT + AY and Pr[YES|NO] = AY. The ratio is thus (AT+AY)/AY = 1.1, and AT = (0.1)AY. That's the same as the slide so far.
      Now I can plug in any value I want (within reason, obviously) for AY, and get the corresponding value for AT and AN. Or I can go the other way. Suppose I devote 0.25 of the spinner to NO. Then I end up with approximately 0.682 for YES and 0.0682 for NO.
      Hope this helps!