PII masking for AI

PII masking for training data, realistic instead of blacked out

For AI training, fine-tuning and test environments: replace real names, addresses and IDs with realistic random fake values. The text stays natural, the people disappear.

  • Realistic fake values
  • Consistent names
  • Bulk via API
  • Irreversible
Microsoft ExcelClient_List_Q3.xlsx
StartEinfügenFormelnDatenAnsicht
A2fxJohn Smith
ABCDEF
1ClientDate of birthIBANEmailPhoneFees
2John SmithMarch 4, 1970GB82 WEST 1234 5698 7654 32j.smith@mail.com(212) 555-1234£4,250.00
3Jane DoeMarch 12, 1985GB33 BUKB 2020 1555 5555 55jane.doe@outlook.com(415) 555-0188£1,980.00
4Acme Construction Ltd–GB94 BARC 1020 1530 0934 59info@acme-build.co.uk(646) 555-0142£12,400.00
5Thomas BakerNov 21, 1962GB29 NWBK 6016 1331 9268 19t.baker@yahoo.com(312) 555-0199£760.00
6Emily ClarkeAug 7, 1979GB60 LOYD 3096 3412 3456 78emily.clarke@proton.me(617) 555-0123£2,315.00
7      
8      
ClientsFeesDeadlines

Why redaction is not enough for training data

  • Models trained on █████ learn bad language.
  • Real personal data in training sets can be regurgitated later.
  • Test databases copied from production are a common source of leaks.

How it works

  1. 1

    Send data via API

    Text or files, synchronous or as a job.

  2. 2

    Fake values, not originals

    Consistent per document: first and last names match, IBANs pass the checksum.

  3. 3

    Train

    No mapping to the original: replacements are random and irreversible.

Example. Before and after.

Original

Mr. John Smith (SSN 123-45-6789) cancelled on March 3. Smith had been a customer since 2019.

Anonymized

Mr. Paul Rivera (SSN 9VW277C5X) cancelled on March 3. Rivera had been a customer since 2019.

FAQ

Questions. Answered briefly.

Try it with your own document

10 pages per month are free. No subscription, no credit card.