Provable Post-Deployment Deterioration Monitoring

ICLR 2025 Conference Submission3810 Authors

Keywords: Deterioration Monitoring, AI Safety, Trustworthy ML, AI for Healthcare, Guardrails for AI

TL;DR: D-PDDM provably monitors model deterioration requiring no training data during deployment, and performs well in real-worlds datasets.

Abstract: Data distribution often changes when deploying a machine learning model into a new environment, but not all shifts degrade model performance, making interventions like retraining unnecessary. This paper addresses model post-deployment deterioration (PDD) monitoring in the context of unlabeled deployment distributions. We formalize unsupervised PDD monitoring within the model disagreement framework where deterioration is detected if an auxiliary model, performing well on training data, shows significant prediction disagreement with the deployed model on test data. We propose D-PDDM, a principled monitoring algorithm achieving low false positive rates under non-deteriorating shifts and provide sample complexity bounds for high true positive rates under deteriorating shifts. Empirical results on both standard benchmark and a real-world large-scale healthcare dataset demonstrate the effectiveness of the framework in addition to its viability as an alert mechanism for existing high-stakes ML pipelines.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3810