Abstract: End devices in end-edge-cloud systems increasingly generate event-driven time-series data, creating a growing need for energy-efficient processing and privacy protection. Federated learning trains models locally and exchanges model updates instead of raw data, which helps address data silos and reduce the direct exposure of private data. However, federated learning methods based on artificial neural networks often struggle to balance energy consumption, latency, and temporal modeling capability when complex time-series tasks are deployed on resource-constrained devices. Federated learning with spiking neural networks offers a possible solution. Spiking neural networks are well suited to event-driven signals and can reduce energy consumption, while the data remain on the devices. However, existing studies still face difficulties in training, client coordination, and trustworthiness. Training spiking neural networks depends on neuronal dynamics across time steps and surrogate-gradient methods. In federated settings, differences among clients and limited computing and communication resources make local temporal optimization more difficult, which may destabilize global training and introduce additional privacy risks. This paper reviews federated learning with spiking neural networks. It first introduces the basic concepts and overall framework, then summarizes the main challenges and reviews representative studies from the perspectives of algorithm design and system mechanisms. The reviewed methods are also compared and analyzed. Finally, this paper summarizes existing benchmarks and applications and discusses possible directions for future research.