2 Commits (e6f0cee896592b25523f00a726f731538f704802)

Author	SHA1	Message	Date
maks-operlejn-ds	4cc4534d81	Data deanonymization (#10093) ... ### Description The feature for pseudonymizing data with ability to retrieve original text (deanonymization) has been implemented. In order to protect private data, such as when querying external APIs (OpenAI), it is worth pseudonymizing sensitive data to maintain full privacy. But then, after the model response, it would be good to have the data in the original form. I implemented the `PresidioReversibleAnonymizer`, which consists of two parts: 1. anonymization - it works the same way as `PresidioAnonymizer`, plus the object itself stores a mapping of made-up values to original ones, for example: ``` { "PERSON": { "<anonymized>": "<original>", "John Doe": "Slim Shady" }, "PHONE_NUMBER": { "111-111-1111": "555-555-5555" } ... } ``` 2. deanonymization - using the mapping described above, it matches fake data with original data and then substitutes it. Between anonymization and deanonymization user can perform different operations, for example, passing the output to LLM. ### Future works - **instance anonymization** - at this point, each occurrence of PII is treated as a separate entity and separately anonymized. Therefore, two occurrences of the name John Doe in the text will be changed to two different names. It is therefore worth introducing support for full instance detection, so that repeated occurrences are treated as a single object. - **better matching and substitution of fake values for real ones** - currently the strategy is based on matching full strings and then substituting them. Due to the indeterminism of language models, it may happen that the value in the answer is slightly changed (e.g. *John Doe* -> *John* or *Main St, New York* -> *New York*) and such a substitution is then no longer possible. Therefore, it is worth adjusting the matching for your needs. - **Q&A with anonymization** - when I'm done writing all the functionality, I thought it would be a cool resource in documentation to write a notebook about retrieval from documents using anonymization. An iterative process, adding new recognizers to fit the data, lessons learned and what to look out for ### Twitter handle @deepsense_ai / @MaksOpp --------- Co-authored-by: MaksOpp <maks.operlejn@gmail.com> Co-authored-by: Bagatur <baskaryan@gmail.com>	1 year ago
maks-operlejn-ds	a8f804a618	Add data anonymizer (#9863) ... ### Description The feature for anonymizing data has been implemented. In order to protect private data, such as when querying external APIs (OpenAI), it is worth pseudonymizing sensitive data to maintain full privacy. Anonynization consists of two steps: 1. **Identification:** Identify all data fields that contain personally identifiable information (PII). 2. **Replacement**: Replace all PIIs with pseudo values or codes that do not reveal any personal information about the individual but can be used for reference. We're not using regular encryption, because the language model won't be able to understand the meaning or context of the encrypted data. We use *Microsoft Presidio* together with *Faker* framework for anonymization purposes because of the wide range of functionalities they provide. The full implementation is available in `PresidioAnonymizer`. ### Future works - **deanonymization** - add the ability to reverse anonymization. For example, the workflow could look like this: `anonymize -> LLMChain -> deanonymize`. By doing this, we will retain anonymity in requests to, for example, OpenAI, and then be able restore the original data. - **instance anonymization** - at this point, each occurrence of PII is treated as a separate entity and separately anonymized. Therefore, two occurrences of the name John Doe in the text will be changed to two different names. It is therefore worth introducing support for full instance detection, so that repeated occurrences are treated as a single object. ### Twitter handle @deepsense_ai / @MaksOpp --------- Co-authored-by: MaksOpp <maks.operlejn@gmail.com> Co-authored-by: Bagatur <baskaryan@gmail.com>	1 year ago