Robust Federated Learning Handles Real-World Client Churn

Federated Learning (FL) offers a way to train models using decentralized, private data on user devices, but its real-world deployment is often hampered by slow model refresh cycles. This slowness makes models unresponsive to rapidly changing user data, crucial for applications like ad targeting or personalized recommendations. The core issues are clients frequently going offline, diverse data distributions across devices, and delays in receiving ground-truth outcomes. A new framework, FeLiX, addresses these challenges by introducing three key innovations. It uses "streaming-aware availability tiers" to quickly identify available clients, "fresh-utility selection" to prioritize valuable updates from devices that can meet deadlines, and "informativeness-aware, delay-robust aggregation" to incorporate late but high-value data without skewing the global model. Experiments show FeLiX dramatically improves efficiency, reducing the time-to-target accuracy by up to 2.37 times and communication bandwidth by 1.30 times compared to existing state-of-the-art FL systems, even in scenarios with low client availability.