The widespread use of machine learning (ML) for advanced applications necessitates assurance solutions to evaluate non-functional properties like fairness, robustness, and privacy to enhance trustworthiness. Assurance has been identified as the preferred method by policymakers, regulators, and industry stakeholders to address this need. However, current assurance solutions are not suitable for nondeterministic ML-based applications.
Challenges
The assurance of ML-based applications is currently more of an art than a science, leading to ad hoc solutions for specific properties like explainability, fairness, and robustness. Despite societal demand, no comprehensive assurance solution for ML exists, and existing assurance solutions cannot be adapted for ML. This stagnation is due to four unresolved challenges:
- Target Definition: Current definitions of targets of assurance evaluations are unsuitable for nondeterministic ML applications and need to evolve.
- Property Definition: There is a lack of rigorous, commonly accepted definitions for ML properties, which complicates property verification.
- Certification Process: Traditional evidence collection methods are inadequate for ML applications, requiring new models that consider training data, processes, and the ML model itself.
- ML Pipelines: Assurance approaches must support the recursive structure of ML applications, addressing robustness and integrity at multiple levels.
The Moon Cloud AI Assurance
To address the above challenges, we need to reshape traditional assurance solutions according to three main aspects: i) the multifactor of ML-based applications behavior (challenge 1), ii) an ML-specific non-functional properties (e.g., fairness, explainability, and robustness) definition (challenge 2), and iii) ML-specific evidence collection models supporting non-functional properties verification at point 2 (challenge 3), iv) MLOps integrated assurance solution to support the model lifecycle including controls on data and training procedure