Auditable AI Pipelines: Logging and Verifiability in ML Workflows

Abstract

The increased reliance on AI in high-stake domains ranging from finance to healthcare to national security has given rise to mounting concerns about the lack of transparency and accountability in ML workflows. Traditional software audit techniques cannot confer sufficient traceability nor verifiability to complex, data-driven AI systems. This work presents a structured auditable AI pipeline framework that implies the embedding of thorough logging and verification units along all stages of the ML cycle. Thus, with the support of provenance in tracking changes and evidence, automated event logging, cryptographic checks by hashes and, optionally, immutability of records through blockchain, it assures operative transparency and forensic reproducibility. We have experimentally shown that through an MLOps implementation, an audit-ready infrastructure, model traceability and regulatory compliance may all be improved when compared to traditional ML environments. The results reassert the urgent need of designing AI pipelines while accounting for auditability as a first-class citizen and present avenues to remedy accountability for enterprise-scale machine learning systems