Custom Vision Data Security (Storage of training and prediction images, Any On-Prem deployment)

Question

I would like to understand the following regarding Azure Custom Vision:

1. How are the training and prediction images stored?
2. How are the stored images secured?
3. Is it possible to request for training images to be deleted after model has been trained?
4. Is it possible to request for prediction images to not be stored?
5. Is there option for on-prem deployment?

The reason for asking is because our company is currently conducting data security review.
Due to the confidentiality of the images, the way the images are handled will affect if Custom Vision will be a viable option for us.

Appreciate if you could provide some clarity for us to make better evaluate the service. Thanks

Answer 1

@Yi Chao Tan There are different components that are involved in using the custom vision service from training to prediction or inference. The high level overview on how the data is processed by custom vision service is available on this page. I'll try to answer all your questions about each scenario that could help you evaluate the service security offerings.

How are the training and prediction images stored?

• The training images are labeled images that are used for the purpose of training the base model. You upload these images when you create a project. The training images are stored in Custom Vision service. Custom Vision stores all of them for the lifetime of the project, if you do not delete them by calling the delete APIs.
• The prediction images passed to run prediction are processed for prediction, you can store them and associate them with the project for further labeling and training. You can also choose not to store them in Custom Vision. If the images aren't used for training for 30 days, they're deleted automatically.

How are the stored images secured?

• Azure Custom Vision automatically encrypts your data when persisted it to the cloud. Custom Vision encryption protects your data and to help you to meet your organizational security and compliance commitments. Data is encrypted and decrypted using FIPS 140-2-compliant 256-bit AES encryption. Encryption and decryption are transparent, meaning encryption and access are managed for you. You can also use customer managed keys along with azure key vault if your organization has a policy to maintain your own keys for encryption. Please see this section on Enterprise readiness that details the other networking options along with key management.

Is it possible to request for training images to be deleted after model has been trained?

• Yes, you can use the delete APIs to delete images, iterations and projects. Please see this page for all the delete APIs available.

Is it possible to request for prediction images to not be stored?

• Yes, see this section on how to call the prediciton API with an option to delete the image immediately after prediction. By default, the images are retained if you want to retrain the model within 30 days. If the model is not retrained the prediction images are automatically deleted.

Is there option for on-prem deployment?

• Yes, you can export your model in the available export formats and use the same for offline prediction. However, the export is done with compact domain and predictions with exported model might be less accurate than an online model.

I hope this helps!!

If an answer is helpful, please click on or upvote which might help other community members reading this thread.