Report on Current Developments in Federated Learning

General Direction of the Field

The field of Federated Learning (FL) is rapidly evolving, with recent advancements focusing on enhancing the efficiency, scalability, and robustness of FL frameworks. A notable trend is the integration of satellite networks into FL, addressing the limitations of terrestrial networks in terms of coverage and bandwidth. This integration aims to leverage the global reach of low-Earth orbit (LEO) satellites to augment traditional FL, particularly in scenarios where ground-based communication is constrained. The challenge of managing heterogeneous environments, where devices vary in data distribution, bandwidth, and computing power, is being addressed through innovative techniques that optimize model aggregation and local training processes.

Another significant development is the emphasis on security and flexibility in FL frameworks, particularly for sensitive data domains such as biomedical research. These frameworks are designed to operate within secure, multi-party consortiums, ensuring that data remains within the consortium's internal network while still enabling collaborative model training. The use of homomorphic encryption and other privacy-preserving techniques is becoming more prevalent, allowing for secure computation on encrypted data.

The field is also witnessing a push towards faster prototyping and deployment of FL systems, especially in vertical federated learning (VFL) scenarios. VFL, where data is partitioned by features across different organizations, requires specialized tools that abstract away the complexities of distributed engineering, allowing researchers to focus on algorithmic innovation. Open-source frameworks are emerging that provide robust support for VFL, enabling rapid development and testing of new algorithms.

Energy efficiency is another critical area of focus, particularly in resource-constrained environments like satellite constellations. Novel scheduling algorithms are being developed to optimize energy usage during model training, extending the operational lifetime of satellites without compromising convergence speed.

Finally, there is a growing recognition of the need to support complex, hierarchical network topologies in FL. Traditional two-tier FL models are being augmented with frameworks that can handle multi-tier, asynchronous aggregation, reducing communication overheads and improving scalability.

Noteworthy Innovations

• SatFed: Introduces a freshness-based model prioritization approach to optimize satellite-ground bandwidth usage, enhancing FL in heterogeneous environments.
• Flotta: Offers a secure and flexible FL framework inspired by Apache Spark, tailored for high-security, multi-party research consortia.
• Stalactite: Provides an open-source toolbox for rapid prototyping of VFL systems, with built-in support for homomorphic encryption.
• Energy-Aware Federated Learning: Proposes a novel energy-aware scheduler for satellite FL, significantly extending battery lifetime without affecting convergence.
• Flight: Introduces a FaaS-based framework for complex, hierarchical FL, reducing communication overheads and improving scalability.