Choosing an approach

Tokenize, encrypt or redact

Three ways to stop holding an identifier, routinely treated as interchangeable and not remotely so. The question that separates them is not how strong each one is. It is what you can still do afterwards, and who can undo it.

Talk to our technical team → — for integration questions, a security review or a volume estimate.

The one question that decides it

Do you need to match records afterwards, and do you ever need the original value back? Those two answers pick the approach; everything else is detail.

Neither
Redact. The value is gone. Nothing to protect, nothing to key management, nothing to breach.
Match, but never recover
One-way tokens. Records still join. The value is unrecoverable by anyone, including the provider.
Match and recover
Vault tokenization. Powerful, and the vault becomes the most sensitive system you operate.
Recover, in place
Encryption. The data stays in your schema and your obligations stay with it.

Most teams reach for encryption because it is familiar, then discover the compliance position has not moved: they still hold the data, and now they also hold a key that must be rotated, escrowed and audited.

Side by side

RedactionOne-way tokenVault tokenEncryption
ReversibleNoNoYes, by the vaultYes, with the key
Records still joinNoYesYesOnly with deterministic mode
You still hold the valueNoNoYes, in the vaultYes, as ciphertext
Breach of that store yieldsNothingNothingEverythingEverything, given the key
Key managementNoneOne server secretFull lifecycleFull lifecycle
Preserves formatNoNoOptionallyOnly with FPE
Safe to send to a third partyYesYesThe token, yesCiphertext, but they cannot use it
Deterministic encryption is not a free lunch. It is what people reach for when they need encrypted values to still join, and it leaks equality: anyone who can see the column learns which rows share a value, and with a low-cardinality field that is often enough to recover the plaintext by frequency alone.

Two kinds of tokenization, and the difference matters

“Tokenization” is sold as one thing and is really two, with opposite risk profiles.

Vault-based
The provider stores the original value and hands back a random token. Reversible, format-preserving if you want it, and the vault now holds every identifier you were trying not to hold. You have moved the risk, which is often the right call — but you have not removed it, and your vendor's breach is now your breach.
Derived
The token is an HMAC of the value under a secret salt. Nothing is stored, so there is no mapping to leak and no vault to audit. Deterministic, so records still join. Irreversible by construction — the provider cannot recover the value either, which is the point and also the limitation.

Ask any tokenization vendor one question: can you return the original value? If yes, they are holding it, and everything that follows from that is now part of your threat model. If no, you cannot recover it either — decide before you tokenize a decade of records.

The mistake that quietly undoes it

If tokens are derived from a single global secret, the same identifier produces the same token for every customer of that provider. Two organisations comparing token lists could then match individuals across their databases without either ever holding the identifier.

That is precisely the property that made the raw identifier dangerous in the first place — a national ID number is harmful mostly because it is a key that joins across otherwise unrelated systems. A globally-salted token rebuilds that key and calls it privacy.

Compliance, briefly and carefully

A general principle, then the caveat that matters more.

Redaction and one-way tokenization remove the value from your systems. Encryption and vault tokenization keep it in a protected form, so obligations that attach to holding personal data generally still apply — and if a third party holds the key or the vault, that relationship is usually in scope too.

Do not take a vendor page as a compliance opinion, this one included. Which approach satisfies GDPR, India's DPDP Act, HIPAA or PCI DSS in your circumstances turns on details no page can know: what you process, why, under which lawful basis, and who else touches it. Every one of these frameworks treats reversibility as the hinge, so start by being precise about who can reverse what.

What this API does

Two of the four, deliberately.

Redaction
The identifier is covered and the page rasterised, so nothing is recoverable underneath. Document in, document out.
One-way tokens
HMAC of the normalised value under a salt derived per key. Same identifier, same token, so records join. Nothing stored, no detokenize call, and no vault to breach.

There is no reversible vault here and that is a decision rather than a gap: a vault would mean holding identifiers, which would contradict the retention position the rest of the product rests on. If you need the value back, you need a different tool, and you should be told that plainly rather than sold this one.

curl -X POST https://api.maskaadhaar.com/api/v1/tokenize \
  -H "X-API-Key: $MASKAADHAAR_KEY" \
  -F "file=@document.pdf" \
  -F "types=global"

{"success": true, "reversible": false,
 "tokens": [{"type": "iban", "token": "iban_a07945e8a890"}]}

Which identifiers it looks for is configurable; the full list and the checksum behind each one is on the identifier reference.

Questions we get asked

What is the difference between tokenization and encryption?

Encryption is reversible by design and the ciphertext is mathematically related to the value, so whoever holds the key holds the data. Tokenization replaces the value with an unrelated substitute; whether it can be reversed depends on whether a mapping is stored. A one-way token has no mapping and cannot be reversed by anyone, including the vendor.

Can tokenization be reversed?

Only if the provider stores a mapping. Vault-based tokenization keeps the original value and can return it, which makes the vault the most sensitive thing in the system. Derived tokens are an HMAC of the value under a secret salt, with nothing stored, and are irreversible by construction.

Does a token still let me match records?

Yes, if the token is deterministic: the same input always produces the same token, so joins and duplicate detection still work. That is usually the reason tokenization is chosen over redaction, which removes the value entirely and takes the join key with it.

Which one satisfies a data minimisation requirement?

Redaction and one-way tokenization both remove the value from your systems. Encryption and vault-based tokenization keep it, in a protected form, so the obligations that attach to holding the data generally still apply. Check the specific framework rather than the general principle, and check who holds the key.

Should the same token appear across different customers?

No. If a token is derived from a single global secret, the same identifier produces the same token for every customer, which turns the token into a cross-organisation correlation key — the exact property that made the raw identifier dangerous. Salting per customer prevents it.

Related

Talk it through with someone technical

Tell us what you need to be able to do with the data afterwards and we will tell you which of these fits — including when the answer is not us.