Protecting sensitive data in AI apps

Рет қаралды 1,649

Google Cloud Tech

Күн бұрын

Пікірлер: 12

@googlecloudtech Ай бұрын

Check out more episodes of Serverless Expeditions with Martin Omander. → goo.gle/ServerlessExpeditions

@eklok5000 Ай бұрын

I know. Only an example. But shouldn’t the image of the person + signature also be redacted in 4:12 ?

@TheMomander Ай бұрын

Good point! The Dataloss Prevention API is focused on preventing loss of sensitive text. For facial detection, you can use Google's Vision API. It can pinpoint faces, tell you if the person is wearing headwear, and if they are smiling, frowning, etc.

@sathish297 Ай бұрын

Thanks Martin for bringing this out. I will give this a try. Is it possible to deidentify before sending to the model and add the personal data back once it is retrieved from the model? Use case would be a school document service where the student information in a document would need to be stripped before sending to the model and added back to the response to the user. ?

@TheMomander Ай бұрын

Good question! Yes, it's called "re-identifying" data in the docs for the Data Loss Prevention API. If you do a web search for "google dlp reidentify quickstart" you'll see a tutorial that walks you through the process step-by-step.

Ай бұрын

There are techniques proposed by the DLP library to do it. For example: Reversible transformations When you de-identify data using the CryptoReplaceFfxFpeConfig or CryptoDeterministicConfig infoType transformations, you can re-identify that data, as long as you have the CryptoKey used to originally de-identify the data. For more information, see Crypto-based tokenization transformations.

@khushahmed97 Ай бұрын

How much latency does this call add ?

@TheMomander Ай бұрын

When I used this API, latency varied a lot with the type and amount of data I sent to it. You can try it out with your data at no cost, as the first 1 GB of data per month is free.