Federated Learning
Federated learning coordinates model updates from distributed clients without centralizing raw data; it couples optimization with device heterogeneity, communication, privacy, participation, and failure behavior.
核心思想
A server selects clients, distributes a model, aggregates returned updates, and repeats. The abstraction does not itself guarantee privacy or robustness: secure aggregation, differential privacy, participant selection, and systems scheduling provide separate properties under stated threat and availability models.
为什么重要
Real cross-device deployments face intermittent connectivity, non-IID data, limited uplink, dropouts, and resource-constrained clients. A protocol speedup or learning result must state whether it is a timing model, simulation, or deployment measurement and which server/client adversaries it covers.
关键观察 / 隐含假设
- 观察：aggregation overhead and privacy requirements can dominate the round. DISAGG-MLSys26 evaluates a committee-based secure-aggregation design under parameterized timing and simulation boundaries.
- 观察：client heterogeneity and participation affect systems policy. PLayer-FL-MLSys26 and FLoRIST-MLSys26 study federated-system design contexts.
- 假设：local training and aggregation improve a common objective despite data and availability variation. AssyLLM-ATC25 and SONAR-MLSys26 expose workload/system-specific boundaries.
设计空间与取舍
- Privacy vs communication/compute：secure aggregation and privacy mechanisms add protocol and endpoint cost.
- Synchronous vs asynchronous participation：asynchrony reduces waiting but changes staleness and correctness assumptions.
- Statistical utility vs systems feasibility：model convergence, data heterogeneity, and device availability need joint evaluation.
引用本概念的论文
- DISAGG-MLSys26 — secure aggregation under a federated threat model.
- PLayer-FL-MLSys26 — federated system design.
- FLoRIST-MLSys26 — federated learning/runtime context.
- AssyLLM-ATC25 — asynchronous/distributed learning boundary.
- SONAR-MLSys26 — federated-system behavior.