The Cross-Industry Standard Process for the development of Machine Learning applications with Quality assurance (CRISP-ML(Q)) framework describes the full AI model lifecycle from data sourcing to deployment, along with risk mitigation procedures for each phase of development. In this work we implement an example model pipeline that utilizes CRISP-ML(Q), along with mitigations for risks documented in MITRE ATLAS and the National Institute of Standards and Technology Artificial Intelligence Risk Management Framework (NIST AI RMF) and its associated playbook, with the goal of creating a generalizable reference framework for model developers and decisionmakers. This work builds on our previous effort in cross-referencing CRISP-ML(Q) with published ML risk frameworks by developing a working and reproducible pipeline for image classification that validates our risk mitigation approach. In this pipeline, we implement multiple risk mitigations and demonstrate their effectiveness in the areas of drift detection, spam mitigation, and defense against adversarial attacks, along with demonstrations of model explainability to improve model trustworthiness. We present results from a variety of A/B mitigation tests that show the effectiveness of our mitigations, and we demonstrate full pipeline runs for Convolutional Neural Network (CNN) and Resnet18 model architectures that utilize our risk mitigation process.
|