Promoting Collaborations in Cross-Silo Federated Learning: Challenges and Opportunities

In cross-silo federated learning (FL), companies or organizations collectively train a shared model while keeping the raw data local. The success of cross-silo FL relies on client cooperation, effective communication, and sufficient resource contributions for model training. However, several unique challenges make client collaboration in cross-silo FL difficult. First, as the global model is a public good, clients may choose to free ride on the process instead of actively contributing to the training process. Second, market competition among clients also discourages their collaborations in training, as clients may not want their business competitors to obtain a high-quality model. Third, repeated interactions among clients may further decentivize collaboration, as one can free ride on others' long-term active contributions. This paper focuses on designing effective economic mechanisms to address the above challenges. Specifically, we propose an incentive mechanism to address the public good issue, a revenue-sharing mechanism to mitigate business competition, and a cooperative strategy to enable clients' longterm collaboration. Our results provide insights into better design of collaboration mechanism and communication in practical cross-silo applications. We further discuss some future directions and open issues that merit research efforts from the community.