Practice: Privacy

Question 0

Consider the following fake dataset showing medical records. The dataset has been fuzzed so that some of the values are not exact. What is the largest value \(k\) such that this dataset provides \(k\)-anonymity? Use the columns “Age”, “Sex”, and “Zip Code” for the insensitive columns that can identify a person.

Note that a * indicates a fuzzed value or part of a value that is a notation to say “any value”.

Name	Age	Sex	Zip Code	Diagnosis
*	20-30	Male	98*	Cardiovascular
*	20-30	Male	98*	Respiratory
*	20-30	Male	98*	None
*	20-30	Female	10*	Cancer
*	20-30	Female	10*	None
*	20-30	Female	10*	Cardiovascular

Your Task

Write your answer down in your own space.

Question 1

Suppose we had a series of \(\varepsilon\)-differntially private algorithms. Order them from the strongest privacy guarantee to the smallest privacy guarantee .

Your Task

Reorder the following options. Write your answer down in your own space.

Option 0

10-differentially private

Option 1

0-differentially private

Option 2

1-differentially private

Option 3

0.5-differentially private

Question 2

Consider our discussion of a differential privacy mechanism for jittering published statistics with the Laplace Mechanism. Select all of the following statements that are generally true.

Your Task

Select one or more options. Write your answer down in your own space.

Option 0

Providing stronger guarantees for privacy requires adding more noise to the result.

Option 1

Providing stronger guarantees for privacy requires adding less noise to the result.

Option 2

Providing stronger guarantees for privacy requires generating noise from a Laplace distribution with a higher \(\varepsilon\) parameter.

Option 3

Providing stronger guarantees for privacy requires generating noise from a Laplace distribution with a lower \(\varepsilon\) parameter.

Option 4

Providing stronger guarantees for privacy requires generating noise from a Laplace distribution that has more likelihood of numbers close to 0.

Option 5

Providing stronger guarantees for privacy requires generating noise from a Laplace distribution that has more likelihood of numbers far to 0.

Question 3

Consider the randomized response differential privacy mechanism. We will use the procedure described in the reading where the respondent flips a fair coin that is equally likely to show up Heads/Tails.

Suppose after calling everyone in our sample, we reported that 9/20 of the respondents answered “Yes” when following the randomized response mechanism. What is our estimate of the fraction of the population that has “Yes” as their true answer (which they could have truthfully told us “Yes” 3/4 of the time, or told us a wrong answer 1/4 of the time). Enter your number as a probability (e.g., 0.33 precise to 2 decimal places).

Your Task

Write your answer down in your own space.