Version: 1.3.14

No Model Access Scans

This page provides an introduction to scans that are configured and run based on prior model predictions and actual outcomes when there is no direct access to the model.

Configuration for scans with no model access may be completed using:

Configuration Specifications

  • In the scan definition evaluation.no_model_access is set to true. (by default it is False).
  • One and only one model is identified in the scan definition.
    • Model metadata is required for Scan Manager to store and manage all use cases and scans. (model_name, model_id, and model_type)
    • The model config may specify a predict_url or a local model, but it is not called during runtime
  • Explanation Type is set to counterfactuals.
  • Report Type selections are restricted.
  • Dataset schema must specify outcome and predicted_outcome columns.
  • Evaluation dataset must specify the predicted_outcome column.

Report Types

No model access scans reports are run using an alternative counterfactual method based on predicted outcomes and ground truth provided in the datasets.

Because the standard Genetic algorithm is not used some scan report types are NOT available for no Model access scans including:

  • Fast Explanations
  • Preflight scans
  • SHAP
  • Fairness - burden-based metrics
  • Robustness

The following reports may be configured for no Model access scans provided the datasets contain the required attributes.

  • Fairness (ground_truth, predicted_outcome)

    • demographic_parity
    • disparate_impact
    • equal_opportunity
    • equal_odds
    • sufficiency
  • Performance (ground_truth, predicted_outcome)

  • Explanations (predicted_outcome in the Explanation dataset)

Counterfactuals for no Model access

Without model access, Certifai is limited in its capabilities to find Counterfactuals explanations because model predictions are not available for data points outside the provided datasets. For this reason, Certifai is unable to run Counterfactual based analyses where the measured score is highly dependent on finding optimal Counterfactuals (i.e. Robustness, Explainability, and burden based Fairness).