Protecting enterprise IP and customer PII

Deployment options for A8Anon

• A8Anon API
• Service hosted in the Autonom8 Google cloud (best price)
• Service hosted in your managed cloud (best data custody)
• Hosted on-prem (contact us for details)

What is Anonymization?

In our context, anonymization is the process by which textual data is cleansed of information (aka PII) which can identify a person, corporation, or transaction. The process first requires identification of PII in text, after which a variety of cleansing treatments can be applied to achieve the desired result. These techniques are briefly described below. A8Anon supports all these capabilities.

Redaction

Redaction is the process of masking PII in text. PII can be removed or obsured. For readability it is sometimes useful to perform partial masking e.g. leaving the last 4 digits of a phone number intact. A8Anon supports a variety of redaction styles including full and partial removal and obscuration with patterns such as LAST4 and ASTERISKS.

Pseudonymization

Pseudonymization replaces PII with false but plausible text. This maintains continuity of textual narrative but disguises identities. It is applicable to names, phone numbers, email and more. In retail situations consumers can be disassociated from their purchases or location. Pseudonymized prompts can be sent to LLMs without risk of identity disclosure. A8Anon enables a private key map for reverse lookup. Responses from LLMs can be mapped back to names which are meaningful to the consumer.

Encryption/Hashing

Encryption is the process of applying a cipher to change text to something which is implausible. Encryption is generally more applicable for numerical data e.g. quantity, price. By definition, encrypted data can be decrypted with a 2-way key. Encryption with a 1-way key is called hashing. Hashed text cannot be recovered. A8Anon supports encryption and hashing for all types of PII.

Generalization (aka Bucketing)

Generalization is an approach for grouping (mostly numerical) data into broader categories to hinder identifiability and thereby enhance privacy. These broader ranges are called buckets. Text such as “He is 43 years old” becomes “He is 40-45 years old”. Replacing a street address with a zip code is another example of generalization.

Additional techniques

• Perturbation: Adding noise to data to disguise numerical values such as salary or dates
• Swapping: Useful for values selected from a list. E.g. If customers are classified as ‘blue’, ‘silver’, and ‘gold’, we can shuffle these values
• Secure user actions: Let users run their own code when certain data classes are detected