Neural Architecture Search in Federated Learning: A Unified Design Concept

While federated learning (FL) enables collaborative model training without centralizing data, it typically relies on predefined neural architectures that may be poorly aligned with decentralized, heterogeneous, and resource-constrained clients. Such architectural mismatches can lead to inefficient computation and communication, degraded model performance, or even infeasible deployment. Neural architecture search (NAS) can address these issues by tailoring neural architectures to different characteristics of FL systems. However, decentralized training renders many conventional NAS methods ineffective. In response, numerous NAS methods tailored to FL have been proposed; yet they differ widely in design and traits and are rarely described in a way that enables systematic comparison. This leaves practitioners with limited guidance on selecting or designing NAS methods for FL systems with different characteristics. Based on a qualitative literature analysis, we conceptualized a generic NAS-in-FL pipeline, identified key design dimensions and associated design elements, and grouped existing NAS methods for FL into five archetypes with distinct traits. Together, these contributions provide a structured foundation for comparing, selecting, and designing NAS methods tailored to FL systems.