Federated learning and information sharing between competitors with different training effectiveness

Abstract

Federated Learning (FL) is an innovative technique that allows multiple firms to collaborate in training machine learning models while preserving data privacy. This is especially important in industries where data is sensitive or subject to regulations like the General Data Protection Regulation (GDPR). Despite its substantial benefits, the adoption of FL in competitive markets faces significant challenges, particularly due to concerns about training effectiveness and price competition. In practice, data from different firms may not be independently and identically distributed (non-IID) and heterogenous, which can lead to differences in model training effectiveness when aggregated through FL. This paper explores how initial product quality, data volume, and training effectiveness affect the formation of FL. We develop a theoretical model to analyze firms’ decisions between adopting machine learning (ML) independently or collaborating through FL. Our results show that when the initial product quality is high, FL can never be formed. Moreover, when the initial product quality is low, and when data volume is low and firms’ training effectiveness differences are small, FL is more likely to form. This is because the competition intensification effect is dominated by the market expansion effect of FL. However, when there is a significant difference in training effectiveness, firms are less likely to adopt FL due to concerns about competitive disadvantage (i.e., the market expansion effect is dominated by the competition intensification effect). This paper contributes to the literature on FL by addressing the strategic decisions firms face in competitive markets and providing insights into how FL designers and policymakers can encourage the formation of FL.