InnoGE Lab · Deckname (v1)
Deckname: anonymize text without it leaving your infrastructure
Deckname detects personal data in German and English text and replaces it with placeholders. The model runs inside your own network, not in a cloud. Paste a text below and see what Deckname finds.
0 / 2.000 characters
Your input is not stored. Processing runs entirely on our own servers in Germany. The text never leaves our infrastructure and is discarded once the response is sent.
Deckname in 32 seconds
What happens when you anonymize a text here
-
Two ways of detecting
IBANs, email addresses and phone numbers are found by a rule layer that recomputes every checksum. Names, locations, professions and organisations are found by the trained model. Hover a highlight and you will see which of the two found it, and with what confidence.
-
Placeholders instead of redaction
The same person gets the same placeholder throughout the text: "Peter Müller" becomes [PERSON_1] everywhere. That keeps the text readable and still usable for a downstream language model. Redaction loses those relationships.
-
Milliseconds, no GPU
Below every result you see the raw model latency. Deckname runs on ordinary CPUs and needs no hardware of its own.
-
This demo is deliberately capped
It runs on our own hardware: up to 2,000 characters per request and a limited number of requests per hour. Neither cap applies when Deckname runs in your environment.
How Deckname text anonymization works
The first model version is technically named InnoGE/deckname (v1). You run it as a container inside your own network. There is no cloud API that text is sent to, and no per-request billing.
PII detection for German and English text
Both languages are covered by the same system. Every replacement is recorded in a machine-readable audit report: which entity was detected at which position, which type it was assigned, and what it was replaced with. For how that is built and what our measurements show, read the Deckname launch article.
Putting the original data back after processing
Nothing is lost. For every replacement, Deckname returns the mapping between placeholder and original value as a structured report: which name sits behind [PERSON_1], which account number behind [IBAN_1], each with its position in the text. That report stays on your systems and does not travel with the text. So you can have the anonymized text processed externally, by a cloud language model for instance, and put the original values back locally afterwards. That mapping is exactly why the process supports your pseudonymisation: the link to the original still exists, but it sits with you alone.
GDPR text anonymization: what Deckname does
Deckname detects and replaces personal data. It thereby supports your anonymization and pseudonymisation processes and can sit in your data protection concept as a technical and organisational measure under Art. 25 and Art. 32 GDPR. Whether a given result counts as anonymous or pseudonymous in law is decided by context: the residual information left in the text, the additional knowledge available to the receiving party, the purpose of processing. Your data protection organisation makes that call.
Let us talk about your use case
This demo runs on our servers. In production, Deckname runs on yours: as a container inside your network, with no data egress. In a free first call we work out which texts you deal with and what running Deckname looks like in your environment.
Benchmark sources: GermEval 2014 NER (Benikova, Biemann, Reznicek) · German LER (Leitner, Rehm, Moreno-Schneider) · License: CC BY 4.0