Canadian Social Insurance Number (SIN) Redaction

What this policy does

Detects and redacts the Canadian Social Insurance Number (SIN), the nine-digit number issued to individuals to work in Canada and access government programs. The SIN appears in onboarding paperwork, payroll, tax forms, and benefits records, so it is a common identifier to strip before sharing or retaining those documents.

The policy uses Phileas’s generic identifier filter with two parts:

• A pattern that matches a nine-digit number written either unformatted (046454286), space-separated (046 454 286), or hyphenated (046-454-286).
• The luhn validator, which keeps a match only if its digits pass the standard mod-10 Luhn checksum. The SIN is a Luhn-valid number, so this rejects the many nine-digit values that merely look like a SIN but are not (an order number, a padded account number, an arbitrary nine-digit reference).

Matches are replaced with [REDACTED-CANADA-SIN].

Why the validator matters

A bare nine-digit pattern over-matches: any run of nine digits would be redacted, including reference numbers, partial account numbers, and other non-SIN values. The Luhn check is what makes this entry precise without a dedicated, code-heavy filter. For example, 046 454 286 is Luhn-valid and is redacted, while 123 456 789 matches the same shape but fails the checksum and is left in place.

The validator narrows false positives; it does not eliminate them. A random nine-digit number passes Luhn roughly one time in ten, so a Luhn-valid value is not guaranteed to be a SIN. For higher precision in free text, anchor on context (see “When to customize”), and always validate the policy against your own representative documents.

When to use this

• Employee and HR records: onboarding forms, payroll files, T4 and tax slips, benefits enrollment.
• Records shared with a third party: payroll processors, benefits administrators, or auditors who do not need the raw SIN.
• De-identifying datasets drawn from Canadian customer or employee data before analytics or model training.

When to customize

• Anchor on context for higher precision. In free-form text where bare nine-digit numbers are common, require a nearby cue and redact only the captured digits with groupNumber:

  {
    "classification": "canada-sin",
    "pattern": "\\b(?:SIN|Social Insurance Number|num[ée]ro d'assurance sociale)[\\s:#-]*((?:\\d{3}[ -]?){2}\\d{3})\\b",
    "caseSensitive": false,
    "groupNumber": 1,
    "validator": "luhn",
    "identifierFilterStrategies": [
      { "strategy": "REDACT", "redactionFormat": "[REDACTED-CANADA-SIN]" }
    ]
  }
  

  This matches only a SIN introduced by “SIN”, “Social Insurance Number”, or the French “numéro d’assurance sociale”, and redacts just the number, leaving the label in place. It trades recall (bare SINs with no nearby cue are missed) for precision.

• Mask instead of redact. To keep the last digits visible for reconciliation, swap the strategy to MASK with a maskLength rather than REDACT.

• Tighten the separators. The default accepts space or hyphen between groups, including mixed separators. If your documents use exactly one convention, narrow the pattern to it.

Prerequisites and compatibility

This policy depends on capabilities added in redaction policy schema 1.1.0:

• The validator field on the identifier filter (added in schema 1.1.0).
• The luhn validator, implemented in Phileas 4.1.0 (Java).

Use Phileas 4.1.0 or later, or a Philter release that bundles it. On an older build that predates schema 1.1.0 or the luhn validator, the validator field is not applied and the identifier would redact every nine-digit match (including checksum-invalid ones), so do not rely on this policy there.

The luhn validator currently ships in Phileas (Java). The phileas-python and phileas-dotnet bindings do not implement it yet, so this policy validates as intended only on the Java runtime until that parity lands. The example input and output below were verified against Phileas 4.1.0.

References

• Social Insurance Number (Government of Canada)
• Luhn algorithm