← Back to blog

Differential Privacy for Shared Analytics Reports

Removing names or hiding small cells does not make an analytics report private. Differential privacy is measurable only when its budget and query surface are designed together.

An aggregate report can reveal information even when it contains no names or account IDs. A small cohort, repeated exports, or two queries that differ by one person's activity can expose whether that person contributed. Treating a dashboard as “anonymous” because it shows totals is not a privacy guarantee.

This problem is newly concrete for teams building shared analytics and public reporting. In March 2025, NIST published SP 800-226, guidance for evaluating differential privacy guarantees. Its publication makes a useful shift for engineering teams: privacy claims should be assessed as properties of a release mechanism, not inferred from the shape of the output.

Aggregation does not erase disclosure risk

Suppose a report shows one visit for a rare campaign, or a customer can filter to a single date, country, and page. The total may effectively describe one visitor or one business. Suppressing small cells helps with some direct disclosures, but an attacker can combine overlapping totals, compare successive snapshots, or use outside knowledge to infer a hidden value.

There is no universal cell-size threshold that makes every dataset safe. Risk depends on what the output reveals, what other information is available, who can query it, and how often the output can change. NIST's de-identification guidance similarly warns that removing identifiers is only one possible technique; organizations need to assess the intended release and its disclosure risks.

What differential privacy changes

Differential privacy (DP) defines a mathematical bound on how much the output of a randomized analysis can change when one person's data is included rather than excluded. It does not promise that no individual can ever be inferred. Instead, it limits the incremental influence of one person's participation under stated assumptions.

The parameter epsilon (ε) is commonly used to express the privacy loss bound. Lower epsilon generally means stronger protection for a given mechanism, but epsilon is not a universal safety score and cannot be interpreted without the neighboring-dataset definition, the mechanism, and composition across releases. A system that spends privacy budget on many queries can reveal more than any single query suggests.

For web analytics, first define what one contribution means. Is it a pageview, a daily visitor, a site, or a user-day? Bound the number and magnitude of contributions from any one person, then design the query and noise mechanism around that bound. If a single visitor can add thousands of rows to a count, adding noise to the final count alone may not give the intended per-person guarantee.

Budget the release surface

Track privacy expenditure across all outputs that draw on the same underlying data. A query builder with arbitrary date ranges, filters, and repeated exports creates a different risk surface from a fixed weekly report. Limit the available dimensions, coarsen rare categories, and decide whether queries are cached, composed, or charged against a finite budget. These are product and API decisions as much as statistical ones.

Make the budget explicit in the system design. Record the protected unit, adjacency model, mechanism, epsilon and any delta, contribution limits, release cadence, and who can initiate a release. When the same report is regenerated, avoid silently treating it as a new independent result. Releasing fresh noise each time can allow averaging; returning a cached result avoids that particular repeated-sampling problem but does not remove all composition concerns.

Do not market a threshold as a guarantee

A minimum cohort size, rounded number, or hidden row is useful risk reduction, not differential privacy. NIST's SP 800-226 is useful because it focuses on evaluating guarantees and their assumptions rather than treating a technique name as proof. NIST SP 800-188 also recommends choosing a data-sharing model, measurable de-identification goals, and risk assessment appropriate to the release.

Not every product needs DP. A private dashboard limited to one authorized customer has a different audience and threat model from a public benchmark or a cross-customer report. But if a system permits repeated queries over many tenants, assume that users can compare outputs and make the privacy model explicit before exposing that flexibility.

Statistical disclosure control is not a cosmetic layer placed on top of a finished dashboard. The unit of protection, query budget, and report design determine whether an aggregate can be safely shared in the first place.

Sources

Comments

Loading comments…