Pseudalis
Use and share your documents without exposing the personal data in them
Pseudalis detects the personal data in your documents and replaces each value with a tag that holds it in encrypted form. Only the key of your computer can restore the document, or any text derived from it.
- Free
- Open source
- Windows 11, x64
- No driver
- Version 1.0.11
- Released 11 November 2026
- License Boost Software License 1.0
Subject: contract renewal Dear Jean Dupont, Following our call of 12/03/2019, please find the renewal for Atelier Dupont et Fils SARL, Saint-Malo. The refund will be made to the account GB29 NWBK 6016 1331 9268 19. You can reach me at 06 12 34 56 78 or jean.dupont@example.com.
Subject: contract renewal Dear [[PERSON:Rnifypn Eouiitlrwky#EKR]], Following our call of [[DATE:28/08/2018#VLC]], please find the renewal for [[ORG:Xkekwmufwmcj Eouiitlrwky et Wldwdst SARL#T4N]], [[LOCATION:Saint-Ntaahgp#QWB]]. The refund will be made to the account [[IBAN:GB45 NWBK 6152 7186 9288 16#K9P]]. You can reach me at [[PHONE:06 07 09 00 95#M2X]] or [[EMAIL:rnifypn.eouiitlrwky@pspbshklkspj.com#7HD]].
Subject: contract renewal Dear Rnifypn Eouiitlrwky, Following our call of 28/08/2018, please find the renewal for Xkekwmufwmcj Eouiitlrwky et Wldwdst SARL, Saint-Ntaahgp. The refund will be made to the account GB45 NWBK 6152 7186 9288 16. You can reach me at 06 07 09 00 95 or rnifypn.eouiitlrwky@pspbshklkspj.com.
Invented values. The substitutes are those of the demonstration key of the research report; the three check characters of each tag are illustrative.
- 0.946
- F1 of the detector on a test set in five languages (version 1.0.11)
- 41,000
- characters read per second on two cores
- 40
- countries whose national numbers are verified
- 17 MB
- installer that carries everything, without driver
The strengths of Pseudalis
Reversible, with a single key
Each value becomes a tag that holds its own ciphertext. With no lookup table, nothing can be lost or stolen: the key of the computer is enough to restore everything.
Documents that remain usable
Word, Excel, PowerPoint and PDF files keep their layout. Substitutes keep the format of the values: exported tables still join and a protected IBAN still passes its check.
Safe prompts for external AI
Pseudonymize a text before sending it to an external model, then restore the values in the answer. With Albireo on the same computer, the relay does it automatically.
No data leaves the computer
Detection, encryption and the vault of originals all run locally. The service listens on the loopback address only.
From detection to restoration
- A tag carries its value
- Each value becomes a tag that holds it encrypted with FF1 over AES-256, with a check. No correspondence table is kept.
- Documents keep their layout
- Word, Excel, PowerPoint, PDF, text, CSV, Markdown, HTML, XML, JSON and YAML files keep their layout, and a PDF keeps its pages.
- With or without tags
- The values can be written as their substitutes alone, so that the file reads as an ordinary document for a prompt, test data or a file sent outside.
- Names and properties
- File names, archive entries and document properties take substitutes too, and come back at the restoration.
- Imposed terms and columns
- Administrators impose up to 100 terms and 100 columns of CSV files, replaced before anything else is read, so that the joins between exported tables hold.
- Originals kept encrypted
- The originals go to an encrypted vault bounded in time and space, and come back on demand from any copy of the anonymized file.
- Pictures described by Albireo
- With Albireo on the same computer, each picture is replaced by a description that names no person.
- Consistent substitutes
- The same value takes the same substitute everywhere, a protected IBAN still passes its check, and the dates move together by a secret shift.
Built into the tools you already use
A drive
A file or a folder copied to the drive of Pseudalis is anonymized once written, its name included. No driver is installed.
Explorer commands
Anonymize and Restore on files, folders and selections of up to 100 items, with or without tags, in place or as a copy.
A web application
Drop files, archives and folders, or paste a text, and get them back anonymized or restored.
A local interface
A documented HTTPS interface on the loopback address for other programs: files, texts, names, records and settings.
Four detection devices
Rules spot the values that have a structure; two small trained models spot, from the context, the ones the rules miss.
Validated scanners
E-mail and IP addresses, telephone numbers, IBAN, cards and the national numbers of about forty countries, accepted when their check digits verify.
Word tables
44 tables in five languages: civilities, field labels, legal forms, given names, public bodies and about two thousand organisations.
Structural SVM tagger
A sequence labeler over words that reads names, places, addresses and organisations from their context.
Byte-level transformer
A span model of 116,791 parameters that reads any spelling and reports where a value starts and ends.
Audiences and uses of Pseudalis
Data protection officers
A concrete pseudonymization and encryption measure for articles 25 and 32 of the GDPR, with its limits documented.
Teams that use generative AI
Prompts and attachments leave without names, addresses or account numbers, and the answers come back complete.
Developers and testers
Production exports become test data that keep their formats and the joins between their tables.
Support and consulting
Documents sent to a supplier or a partner carry substitutes, while the originals stay in the encrypted vault.
Characteristics
| Platform | Windows 11, x64, without driver |
|---|---|
| Documents | Word, Excel and PowerPoint (.docx, .xlsx, .pptx and their macro and template variants), PDF |
| Texts | More than thirty extensions, among them .txt, .csv, .md, .json, .xml, .html, .yaml and .sql, and any file whose content is text |
| Pictures and archives | PNG, JPEG, GIF, BMP, TIFF and WebP pictures; ZIP, TAR and TAR.GZ archives; whole folders |
| Languages read | French, English, Italian, Spanish and German |
| Interface languages | English, French, German, Italian and Spanish |
| Encryption of the values | FF1 over AES-256, each tag authenticated with HMAC-SHA-256 |
| Vault of the originals | AES-256-CTR with HMAC-SHA-256; master key protected by DPAPI; keys derived with HKDF-SHA-256 |
| Export of the key | Passphrase of 12 characters or more, PBKDF2-SHA-256 with 600,000 iterations |
| Local interface | https://127.0.0.1:8444, documented with Swagger |
| Updates | Feed signed with ECDSA P-256, installer checked by size and SHA-256 |
| Dependencies | None: the installer carries everything, and the programs import the libraries of Windows alone |
| License | Boost Software License 1.0, Community edition, with no copyleft component |
Measures of the detector
The figures concern the whole detector, every category, with a prediction counted only when its type and its first and last bytes are exactly right.
| Test set | Precision | Recall | F1 |
|---|---|---|---|
| Test, five languages | 0.955 | 0.935 | 0.945 |
| Values never seen in training | 0.951 | 0.935 | 0.943 |
| Six times the noise | 0.941 | 0.897 | 0.918 |
| Whole documents | 0.961 | 0.943 | 0.952 |
| Places abroad only | 0.960 | 0.939 | 0.949 |
| System | F1 on the test set |
|---|---|
| Rules alone | 0.840 |
| SVM tagger alone | 0.749 |
| Byte-level transformer alone | 0.903 |
| Whole detector | 0.945 |
| By language | Precision | Recall | F1 |
|---|---|---|---|
| French | 0.975 | 0.965 | 0.970 |
| English | 0.971 | 0.964 | 0.967 |
| Italian | 0.950 | 0.926 | 0.938 |
| German | 0.949 | 0.926 | 0.937 |
| Spanish | 0.928 | 0.898 | 0.913 |
| Document | Size | Time |
|---|---|---|
| Text of 5,000 words with 720 values, .txt | 27.5 KB | 2.5 – 2.6 s |
| The same text, .docx | 27.5 KB | 1.4 – 1.5 s |
| CSV of 1,000 rows with 5,027 values | 95 KB | 5.8 – 5.9 s |
Version 1.0.11 reaches an F1 of 0.946 on the test set, 0.919 with six times the noise, 0.953 on whole documents and 0.951 on places abroad; the tables give the detailed measures of version 1.0.5. On hand-written texts in the five languages, the F1 lies between 0.886 and 0.907. On sixteen files written for the test in the five languages (Word, PowerPoint, PDF, CSV, source code, images), version 1.0.11 leaves readable 2 of the 1,223 values to protect, against 109 for the previous version. On a private sample of real documents, precision is 0.859 and recall 0.912: about one value in eleven stays readable, which is why a protected document should be read again before it is shared. The detector reads about 41,000 characters per second on two cores.
Coverage of the GDPR
The protection is a pseudonymization in the sense of the GDPR: the key restores the data, which remain personal data for its holder. Pseudalis provides several of the measures that Articles 25 and 32 name: processing on the computer, pseudonymization and encryption, an encrypted vault with limits in time and space, a function to forget, and signed updates.
| Category | State | Means |
|---|---|---|
| Names, nicknames, logins | Covered | Rules, labels, propagation, models |
| Addresses, cities, postal codes with their city | Covered | Rules, labels, models |
| Telephone, e-mail, dates | Covered | Scanners; dates shifted together |
| Civility, sex, nationality | Covered | Rules, labels |
| Identifiers with a check: social security, IBAN, cards, about 40 countries | Covered | Validated scanners |
| IP, MAC, UUID, IMEI, plates, VIN, GPS | Covered | Validated scanners |
| Organisations and named public bodies | Covered | Rules, tables, models, imposed terms |
| Identity card, driving license, residence permit | Partly | Found through their label |
| Customer, case, contract and employee numbers | Partly | Label, cue or pattern of the organisation |
| Health, beliefs or criminal facts written in whole sentences | Not covered | Left readable |
| Text inside pictures | Not covered | No OCR: the picture is removed |
| NIST SP 800-38G | FF1 format-preserving encryption of the values |
|---|---|
| AES-256, HMAC-SHA-256 | Encryption and authentication of the tags and of the vault |
| HKDF-SHA-256, PBKDF2-SHA-256 | Derivation of the keys and protection of the exported key |
| ECDSA P-256 | Signature of the update feed |
| GDPR, Articles 25 and 32 | Pseudonymization and encryption as measures of protection by design and of security |
| EDPB Guidelines 01/2025 | Pseudonymisation, read as they stood in November 2026 |
The limits of Pseudalis
- Detection is probabilistic in the learned categories: a protected document should be read again before it is shared.
- The trust model is the computer: every user of the computer shares its key, its originals and its administration.
- A scanned page without text layer is kept as a gray box: no OCR is performed.
- A document in a language other than the five that the recognizers read keeps more values readable.
Good to know
- Price and licence
- Pseudalis Community is free under the Boost Software License 1.0, with no copyleft component.
- Formats
- Word, Excel, PowerPoint, PDF, more than thirty text formats, images, ZIP and TAR archives, and whole folders.
- Languages read
- French, English, Italian, Spanish and German. A document in another language keeps more readable values.
- Accuracy
- On its five-language test set, the detector of version 1.0.11, whose models were trained again, reaches an F1 of 0.946. Detection is probabilistic: read a protected document again before sharing it.
- Key
- The key is created on the computer and protected by Windows. It can be exported under a passphrase of 12 characters or more.
Installation
The installer contains the programs, the service, the tray icon and the Explorer commands. Run again over an existing installation, it updates it and keeps the data, the key and the options.
pseudalis-setup /silent [/dir <folder>] [/data-dir <folder>] [/drive <letter>|none] [/port <port>] [/no-context-menu]
pseudalis-setup /uninstall [/silent] [/purge]
pseudalis-setup /service start|stop|restart
pseudalis protect [--copy] [--no-tags] <file or folder>...
pseudalis restore [--copy] <file or folder>...With Albireo on the same computer
Albireo describes each picture of a document in the language of the installation, and Pseudalis reads the description with its detector before writing it. In the other direction, Albireo sends its relayed texts through Pseudalis before they leave the computer.
Install Pseudalis and protect your first document
Copy a file to the Pseudalis drive, or right-click it in Explorer: its anonymized version is ready within moments.