Secure AI

3. Introduction to differential privacy

company not want to share date, and scientists face the challenge.

1. Legal issue to share

2. to maintain competitive advantage

Differential Privacy

1. learns data only supposed to be learned.

2. not learn data not supposed to be learned.

Privacy is preserved

Dwork = a promise: you will not be affected by allowing your data to be used in the study.

If we remove a person from db, and the query doesn't change. => protected.

4. Evaluating the privacy

Sensitivity

• Maximum changes of a query when an individual is removed.

• Maximum difference between fdb query & pdb query.

depends on

1. query functions

2. dataset distribution = removing one row causes sensitivity changes or not.

Differencing Attack

fdb query substract a pdb querry -> reveal the value of the removed part!!!