Atlas: A Framework for ML Lifecycle Provenance & Transparency

We propose Atlas, a framework that enables fully attestable ML pipelines to address ML supply chain risks. Atlas leverages runtime pipeline monitoring and open specifications for data and software provenance to collect model artifact integrity and end-to-end lineage metadata. Atlas combines trusted hardware and transparency logs to enhance metadata integrity and enable efficient verification of ML pipeline operations, from training through deployment. Our prototype implementation of Atlas integrates open-source tools to build an ML lifecycle transparency framework.