If you train a model on DetaBord it won’t be released automatically. Rather it can first be checked by DetaBord‘s privacy checker. In case of a successful privacy check the user can request a model release. Only after the model release the consumed budget is deducted from the used data. Refer to the approval section for more details.

Model Details

You can find information about the trained model in the run details page. As described in run details section you can inspect the model metrics in the “Display” tab.

Go to the “Files” tab to browse the model files. You can inspect additional model information or – if approved – download the model files:

Model 2

Privacy Checker

If you have “data owner” user privileges you can run the DetaBord Privacy Checker to learn more about the data protection guarantees of the trained user model. Select a model file or the model folder in the file browser and click the “Run Privacy Checker” button below the file browser:

Run Privacy Checker

Before you run the privacy checker you need to select and adjust the configuration. Select a template in the lower left and click “Apply” to start with a check for classification or regression models. Also, set the path to the user model to check in the “Module Path” box. This path is relative to the user commit (inspect the “Commit” run details tab for details) folder.

Privacy Checker Configuration

Adjust the privacy checker configuration by editing the yaml properties. You can define the number of runs and configure which tests (attacks) shall be performed.

Click “Run Privacy Checker” to start the privacy checker run. The results of this job will again be inspectable in the corresponding Run Details page. The Privacy Checker run will use the same Experiment as the run of the model that is being checked. Navigate back to the “Runs” page to go to the newly started privacy checker run.

In the Privacy Checker run’s “Files” tab use the file browser to inspect the different attack and checker results:

Privacy Checker Results
Privacy Checker Results

Join Waitlist

Join the AI for Life Sciences Challenge

Let's talk