Privacy-Utility Trade-off Vocabulary Exercises

1 / 5

A data scientist asks: "How synthetic is enough?" when evaluating whether synthetic data can replace real data in a model training pipeline.

What does this question fundamentally address?

"How synthetic is enough?" — the central question of the privacy-utility trade-off:

Dimension	Meaning
Fidelity	How closely the synthetic data mirrors the real data's distributions and correlations
Privacy	How difficult it is to re-identify or infer information about real individuals
Utility	Whether models trained on synthetic data perform as well as those trained on real data

Key vocabulary: privacy-utility trade-off, fidelity, downstream utility, re-identification risk, TSTR (Train on Synthetic, Test on Real).

2 / 5

A privacy engineer explains: "Increasing fidelity of synthetic data generally reduces privacy protection."

Which statement best describes the relationship between fidelity and privacy in synthetic data?

Fidelity vs. Privacy tension:

High fidelity	Low fidelity (more noise)
Synthetic data closely mirrors real distributions	Synthetic data diverges from real data
Higher utility for downstream tasks	Lower utility, noisier models
Higher re-identification risk	Stronger privacy protection

Key vocabulary: membership inference attack, attribute inference, re-identification risk, singling out, linkability, statistical disclosure.

3 / 5

A paper evaluates synthetic data quality using statistical similarity metrics, including the Jensen-Shannon Divergence (JSD) and the Kolmogorov-Smirnov (KS) test.

What do these metrics measure in the context of synthetic data evaluation?

Statistical similarity metrics:

Metric	Measures	Ideal value
JSD	Symmetric divergence between two probability distributions	0 (identical distributions)
KS test	Maximum difference between two CDFs for continuous features	Low statistic, high p-value
Wasserstein distance	"Earth mover's distance" between distributions	0

Key vocabulary: distributional similarity, column-wise statistics, pairwise correlation, marginal distribution, joint distribution.

4 / 5

A data governance team discusses using k-anonymity as a privacy guarantee for a masked dataset that will be shared with researchers.

What does k-anonymity mean?

k-anonymity — a foundational privacy model:

Concept	Explanation
Quasi-identifiers	Attributes that, combined, can single out an individual (e.g., age + postcode + gender)
k-anonymous group	Group of ≥k records sharing the same quasi-identifier values
Limitations	Vulnerable to l-diversity and t-closeness attacks; doesn't protect sensitive attribute values

Related vocabulary: l-diversity, t-closeness, generalisation, suppression, quasi-identifier, re-identification.

5 / 5

A team presents a privacy-utility curve to stakeholders to justify the choice of epsilon for their DP synthetic data pipeline.

How should a privacy engineer interpret a point on this curve where utility is high but privacy is low?

Privacy-utility curve interpretation:

Curve region	Meaning	Suitable for
High utility, low privacy	Data is statistically close to real; re-ID risk is high	Internal analysis only
Balanced (ε ≈ 1–3)	Moderate utility and privacy	Controlled sharing
Low utility, high privacy	Strong privacy guarantee; data may be too noisy	Public release

Key vocabulary: privacy-utility curve, Pareto frontier, epsilon selection, membership inference attack, attribute inference, privacy budget allocation.