Fairness Feedback Loops: Training on Synthetic Data Amplifies Bias

Sierra Calanda Wyllie, Ilia Shumailov, Nicolas Papernot

Primary Area: societal considerations including fairness, safety, privacy

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.

Keywords: fairness, algorithmic reparation, model collapse, trustworthy machine learning, generative models

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We study model-induced distribution shifts and their relationship to unfair feedback loops which we mitigate with algorithmic reparation.

Abstract:

Model-induced distribution shifts (MIDS) occur as previous model outputs pollute new model training sets over generations of models.
This is known as model collapse in the case of generative models, and performative prediction or unfairness feedback loops for supervised models. We provide a taxonomy for MIDS and demonstrate that their long-run fairness effects lead to a lack of representation and performance on minoritized groups within a few generations. We improve upon this unfairness behavior by situating Algorithmic Reparation as an intentional MIDS with the goal of providing redress for historical discrimination and improving the fairness of models subject to other MIDS. Our work makes an important step towards identifying and mitigating the justification of unfair feedback loops via the algorithmic objectivity and idealism ascribed to autonomous systems.

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: 6793