Incentivized Black-Box Model Sharing

Wenyang Hu, Xinyi Xu, See-Kiong Ng, Bryan Kian Hsiang Low

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: incentives, ensemble distillation, collaborative learning, model sharing

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

TL;DR: This paper presents a novel incentivized black-box model sharing framework that fairly distributes rewards and monetary payoffs to each party, and satisfies individual rationality regarding model performance.

Abstract: Black-box model sharing is a preferable alternative to data sharing because of practical considerations (e.g., administrative regulation and data expiration). However, previous works may neglect the self-interests of individual parties. To encourage self-interested parties to contribute predictions in the ensemble, it is crucial to provide incentives, such as __fairness__: allocating higher reward/payoff to parties with more contributions, and __individual rationality__: ensuring guaranteed model performance improvement for each party. This paper presents a novel incentivized black-box model sharing framework that fairly distributes ensemble predictions and monetary payoffs commensurate to each party's contribution. We propose a contribution measure using the average ensemble weight of black-box models. Subsequently, we derive a closed-form solution that explicitly determines the fair reward and payoff allocation given the contribution and payment. By incorporating ensemble predictions and analyzing the generalization error bound, we theoretically show approximate individual rationality is guaranteed. Furthermore, we empirically demonstrate our proposed method achieves incentive guarantee using real-world datasets.

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: 4374