Apply differential privacy to user behavior queries and explain epsilon trade-offs

What's really being asked

This question tests whether you understand differential privacy as a mathematical framework rather than a vague concept. The interviewer wants to see that you know how noise is calibrated, why sensitivity matters, and how the privacy budget epsilon creates a hard constraint that degrades gracefully into statistical uncertainty. Senior candidates should demonstrate they can operationalize DP in a real query system.

The full answer

First, sensitivity analysis. You must explain that the noise scale depends on how much one individual can change the query result. For a count query the sensitivity is one; for a sum it is the maximum possible value. Second, noise injection. You add noise drawn from a Laplace distribution with scale sensitivity over epsilon for pure epsilon-DP, or Gaussian noise for approximate DP. Third, privacy budget accounting. Every query consumes epsilon from a global budget via composition theorems; once exhausted, no more queries are allowed or the guarantee breaks. Fourth, the epsilon-accuracy trade-off. Epsilon is the privacy loss parameter. Typical values in practice range from 0.1 to 10, with smaller values meaning stricter privacy. As epsilon shrinks, the standard deviation of the noise grows linearly, so confidence intervals widen and point estimates become less reliable. Fifth, system design. You would expose a limited query API, precompute sensitivity bounds, and possibly use mechanisms like sparse vector or private aggregation to optimize budget use.

The mistakes people make

Confusing differential privacy with anonymization, pseudonymization, or k-anonymity. Claiming that removing names or user IDs is sufficient. Proposing to add a fixed amount of noise regardless of query sensitivity. Ignoring composition and allowing unlimited queries on the same dataset. Suggesting that epsilon is just a confidence level like in statistics. Another red flag is saying noise makes the data useless without quantifying the error bounds.

What usually comes next

How would you handle a query with unbounded sensitivity like average purchase amount? What happens if researchers need interactive queries versus a one-time release? How do you allocate budget across a team of researchers? Can you explain the difference between pure epsilon-DP and epsilon-delta approximate DP? How would you communicate confidence intervals to researchers who are not privacy experts?

A concrete example

Suppose researchers want to know how many users clicked a button. The true count is 50000. The sensitivity of a count query is 1 because adding or removing one user changes the count by at most 1. With epsilon 1.0, you add Laplace noise with scale 1.0, giving a standard deviation of about 1.41. The reported answer might be 50001 or 49998, which is highly accurate. If you tighten epsilon to 0.1, the scale becomes 10 and the standard deviation becomes about 14.1. Now the result could be 50012 or 49985, still useful at aggregate scale but no longer precise enough to infer any single user's action. If researchers ask 100 such questions, basic composition consumes 100 epsilon total, so you might instead use advanced composition to keep the total budget at perhaps 10 while each query uses a smaller per-query epsilon.