Incentivizing inclusive contributions in model sharing markets.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Public data is nearing exhaustion for training AI models.
• Decentralized private data offers vast potential but faces privacy and incentive barriers.

Purpose of the Study:

• To propose an inclusive and incentivized personalized federated learning (iPFL) framework.
• To enable collaborative training of personalized AI models without raw data exposure.
• To address privacy concerns and incentivize data sharing from diverse holders.

Main Methods:

• Developed a model-sharing market using graph-based training optimization.
• Incorporated a game theory-based incentive mechanism.
• Ensured adherence to individual rationality and incentive compatibility properties.

Main Results:

• iPFL achieved the highest economic utility across eleven AI tasks.
• Demonstrated comparable or superior model performance against baseline methods.
• Empirical studies validated the framework's effectiveness, including for large language models.

Conclusions:

• iPFL offers a viable solution for leveraging private data in AI.
• The framework successfully balances privacy, incentives, and model performance.
• iPFL paves the way for more inclusive and efficient AI development.