New Mechanism Ensures Truthfulness and Fairness in Collaborative Bayesian Learning
A novel approach to collaborative machine learning promotes fairness and honesty in data sharing for Bayesian models. Current data valuation techniques adequately compensate sources according to the data they provide but fail to confirm its authenticity, which opens the door to manipulation via duplication or noise. The suggested mechanism merges semivalues, such as the Shapley value, to ensure fairness with a truthful data valuation function (DVF) that relies on a validation set that sources cannot access. This method guarantees collaborative fairness and encourages honesty in equilibrium. Furthermore, an additional criterion considers the impact of others' data on semivalues, ensuring that sources can only optimize their rewards by submitting accurate data.
Key facts
- Mechanism ensures collaborative fairness and incentivizes truthfulness for Bayesian models
- Combines semivalues (e.g., Shapley value) with a truthful data valuation function
- Truthful DVF based on a validation set unknown to sources
- Existing methods do not verify or incentivize data truthfulness
- Sources can manipulate data (duplicated or noisy) to increase rewards
- First mechanism to provably ensure both properties at equilibrium
- Additional condition introduced to handle influence of others' data on semivalues
- Paper published on arXiv with ID 2605.11889
Entities
Institutions
- arXiv