Disclosure-Risk Review of Released Results

Description

Before sharing aggregates, patterns or segments, check whether any could single out a person or be combined to reveal private values, then suppress, coarsen or perturb them.

Harmless-looking outputs can identify people. A rule or segment covering one customer reveals that person's behaviour, and two permitted totals can be differenced to expose an individual value (total salary minus everyone except the president gives the president's salary). Statistical agencies defend against this with minimum cell sizes, ranges instead of exact values, restrictions on overlapping queries, perturbation of data or outputs, and release of small samples so fine-grained results are unreliable. The review treats every output as a potential inference channel and checks coverage, combinability and proxy discrimination before release. In UX it applies to research repositories, analytics dashboards, segment reports and quotes, where small segments and rich filters make re-identification easy.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. List every output to be released: tables, segment profiles, mined rules, dashboards and quotes.
  2. Check each output's coverage and flag any that describe very few people; census practice uses minimum group sizes, for example releases covering at least 300 people.
  3. Check whether outputs can be combined or differenced (overlapping filters, totals with and without a subgroup) to reveal an individual.
  4. Apply defences: minimum cell sizes, ranges or banding, suppression, perturbation or release of a sample only.
  5. Check use restrictions and whether released variables act as proxies that could lead to unfair treatment.
  6. Record what was suppressed or altered and why, and re-review when new outputs are added.

Sources

  • Handbook of Data Mining (Ye, 2003)
  • The SAGE Handbook of Survey Methodology (Wolf et al., 2016)

Also known as

  • Disclosure-risk review of mined results before release
  • Output disclosure control
  • Inference control for mined patterns
  • Statistical disclosure control

Categories

  • Research Operations
  • Research Design & Rigor