Bridging the Gap: Challenges and Limitations of Federated Learning in Real-World Applications

Federated Learning (FL) has emerged as a key enabler of Machine Learning (ML) applications due its ability to enhance privacy, enable collaboration, improve scalability and possibly reduce environmental impact, all while maintaining high predictive performance comparable to conventional centralized approaches. At the same time, however, FL faces several critical challenges, including statistical heterogeneity, ensuring privacy, personalization and communication efficiency. Although there has been significant research conducted to address these issues, the practical application of these solutions, especially in terms of technical feasibility, remains highly unexplored. For instance, while novel aggregation algorithms have been developed to tackle statistical heterogeneity, they introduce tunable hyper-parameters that may not be practical to adjust in real-world settings. This paper provides a comprehensive overview of the practical issues associated with the open challenges of FL, with a particular emphasis on the practicability of the proposed solutions. We critically evaluate the limitations of these solutions, focusing mainly on federated optimization and privacy-preservation aspects. We outline directions for responsible FL research as an effort to unlock the full potential of FL applications.