Mechanisms for Honest Data Sharing under Competition
summary
NeurIPS 2023 work proposes a concrete mechanism: players send empirical means (k samples each), attackers add Gaussian noise with variance α, and a payment penalizes the squared distance between one player’s message and the mean of others; with penalty constant C ≈ O(1/N^2) honesty (α=0) is a dominant strategy and for N>2 redistributed payments make participation individually rational. The paper proves failure modes for algorithmic defenses, offers a payment-based peer prediction implementation for mean estimation, a noise-reflection personalized replies variant to avoid side payments, extends to federated SGD under strong convexity, and flags heterogeneity and fairness as open problems.
AI-written summary · about
chapters (7)
- 0:00Data sharing challenges in AI coding
- 6:01Firms resist honest sharing under competition
- 11:47Collaborative learning with incentives
- 21:23Honesty penalties vs participation
- 30:19Generalizing data sharing mechanisms
- 43:01Open problems and future research
- 49:02Q&A on peer prediction
related talks
47mNot all data are created Equal: Robust Mean Estimation from Heterogeneous and Unreliable Users
01h 00mTutorial: Incentives for Collaborative Learning and Data Sharing, Part Ib
37mSelfish Robustness and Equilibria in Multi-Player Bandits
27mThe ethical algorithm
28mPrivacy versus Robustness in Federated Learning: Limits and Algorithms
33mPrivacy of Decentralized Machine Learning
01h 27mTutorial: Incentives for Collaborative Learning and Data Sharing, Part II
45mTutorial: Incentives for Collaborative Learning and Data Sharing, Part I