New mechanism tackles false-name manipulation in ML data attribution

Researchers have introduced quotient semivalues as a novel mechanism to address false-name manipulation in machine learning data attribution. This method aims to provide more accurate valuations by clustering data and using a representative operator to mitigate issues like dataset splitting and duplication. The proposed mechanism is designed to be false-name-proof under specific conditions, offering bounded fairness and manipulation loss even with imperfect data provenance. Initial tests in the DataMarket-Gym benchmark show a significant reduction in manipulation gain compared to traditional Shapley values. AI

IMPACT Introduces a new method for fair data valuation in ML, potentially improving the integrity of training data attribution and reducing manipulation.

RANK_REASON Academic paper introducing a new technical mechanism for data attribution. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Banzhaf
• Beta
• DataMarket-Gym
• Florian Burnat
• Quotient Semivalues
• Shapley Values
• Sybil

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →