Welcome to the FLSys Workshop co-located with MLSys 2023 conference.

Federated learning (FL) is an ecosystem of computational approaches and techniques that enable geographically distributed data sources to form secure and private coalitions and jointly analyze their siloed data. Data never leave the original source, and each source only shares its locally trained AI/ML model with the outside world. Through this collaborative learning approach, sources can extract knowledge that is statistically more powerful, less biased, and more accurate when compared to the learning conducted independently by each source. At its core, federated learning uses distributed machine and deep learning methods, and privacy-enhancing technologies to empower learning in a collaborative setting.

There are still many open problems that need to be addressed before federated learning can be widely adopted and deployed in real-world settings. When applied to edge devices, such as Android phones, Raspberry Pi's, and the like, federated learning is known as On-Device FL. Edge devices participating in FL typically experience limitations in computation power, memory capacity, and communication bandwidth. Hardware resources can also vary greatly among such devices, e.g., clients using Jetson Nano with embedded GPUs are able to conduct computations much faster than those with Raspberry Pi equipped with only a quad-core CPU. Edge devices powered by batteries responsible for wireless communication, can suffer from network disconnections, packet drops, or battery drain.

Deploying FL algorithms directly on these edge devices is not feasible without a system that is highly efficient, aware of hardware heterogeneity, and fault-tolerance. Similarly, running extremely large-sized models in a datacenter setting (Cross-Silo FL) over hundreds of clients requires scalable systems that can provide efficient model aggregation and learning task delegation, while at the same time providing failover capabilities and being resilient against numerous attacks (e.g., model poisoning, data poisoning) threatening the federated learning environment.