Cooperate to Compete: Strategic Data Generation and Incentivization Framework for Coopetitive Cross-Silo Federated Learning

A team of researchers has proposed CoCoGen+, a novel framework designed to solve the fundamental 'coopetition' problem in cross-silo federated learning (CFL). In sensitive fields like healthcare, CFL allows organizations to collaboratively train AI models by sharing model updates, not raw data. However, these organizations are often market competitors. The core dilemma is that contributing high-quality, non-IID (non-identically distributed) data improves the shared global model but can inadvertently boost a rival's capabilities, creating a disincentive to participate fully. Existing incentive mechanisms fail to account for this competitive cost.

CoCoGen+ addresses this by framing each training round as a weighted potential game. Organizations strategically decide how much synthetic data to generate using GenAI, balancing the computational cost and the risk of empowering competitors against the performance gains for their own operations. The framework then employs a payoff-redistribution-based incentive mechanism to compensate participants for any utility lost due to competition, promoting long-term, sustainable collaboration. Experiments across various learning tasks demonstrate that CoCoGen+ effectively models how factors like data skew and competition intensity shape organizational strategy, leading to higher social welfare and efficiency compared to previous baseline methods.