Utility-Privacy Trade-off Language Exercises

1 / 5

A privacy engineer tells a product manager: "We accept a 12% utility loss to achieve ε = 1 differential privacy."

What does utility loss mean in the context of privacy-utility trade-offs?

Utility loss = measurable degradation in downstream task performance caused by privacy mechanisms:

Privacy mechanism	Source of utility loss
Differential privacy (noise)	Added noise obscures signal, degrades model accuracy
k-anonymity / generalisation	Coarsened values reduce feature precision
Data suppression	Removed records reduce sample size

Key vocabulary: utility-privacy trade-off, "acceptable utility loss", privacy cost, privacy budget allocation.

2 / 5

A team refers to a privacy-utility curve when deciding on an epsilon value.

What does a privacy-utility curve show?

Privacy-utility curve — decision-making tool:

Region of curve	Interpretation
High ε (right)	Weak privacy, high utility
Low ε (left)	Strong privacy, low utility
Knee of curve	Optimal operating point — best privacy for acceptable utility

Key vocabulary: Pareto frontier, operating point, knee of the curve, "privacy-utility sweet spot", stakeholder requirement alignment.

3 / 5

A data scientist asks: "How synthetic is synthetic enough?"

In practice, what question does this phrase refer to?

"How synthetic is synthetic enough?" — a legal and regulatory question:

Framework	Relevant question
GDPR (EU)	Is re-identification risk low enough to classify as anonymous data?
UK ICO	Does the synthetic data count as personal data under the DPA 2018?
HIPAA (US)	Does the synthetic data qualify as de-identified health information?

Key vocabulary: re-identification risk threshold, "functionally anonymous", singling-out test, ICO anonymisation guidance, privacy risk assessment.

4 / 5

A privacy engineer discusses re-identification risk with a stakeholder.

What is re-identification risk in the context of synthetic data?

Re-identification risk for synthetic data:

Attack vector	How it works
Nearest-neighbour matching	Synthetic record is too close to a real record in feature space
Attribute inference	Adversary infers sensitive attribute of a known individual
Membership inference	Adversary detects if a real record was in the training data

Key vocabulary: distance to closest record (DCR), nearest neighbour adversarial accuracy (NNAA), singling out, linkability, inferability.

5 / 5

An engineer explains to a stakeholder: "We allocated the privacy budget across three queries using composition."

What does privacy budget allocation via composition mean?

Composition in differential privacy:

Composition type	Total epsilon	Note
Basic composition	ε₁ + ε₂ + ε₃	Simple but loose bound
Advanced composition	O(√k · ε)	Tighter for many queries (Dwork et al.)
Rényi DP / zCDP	Even tighter	Used in TensorFlow Privacy

Key vocabulary: privacy budget accounting, budget exhaustion, moments accountant, sequential composition, parallel composition.