As the Internet of Things (IoT) continues to grow, the majority of data generated is being collected from the network edge, which is too burdensome to be sent to a central server for model training due to bandwidth, latency, energy, and most importantly privacy and regulatory considerations. One alternative is federated learning (FL), where devices use raw data to train a shared model without the data ever being sent away from the device. But when deployed at the edge, FL has to face the limits of resource-constrained hardware, unreliable networks, and especially, not being private by default: the model updates that devices exchange can expose a lot of information about the underlying data. This survey provides an overview of the architectures and mechanisms that enable FL to be privacy preserving in IoT environments. It outlines the field along four axes: taxonomy of aggregation topologies and privacy strategies; accuracy, communication, and energy trade-offs of the privacy enhancing mechanisms (differential privacy, secure aggregation, homomorphic encryption, secure multiparty computation, and trusted execution); threat models and attacks that these mechanisms need to withstand (gradient inversion, membership inference, poisoning attacks); and a layered reference architecture spanning the device, edge, and cloud tier, with dimensions along which this architecture is evaluated. The survey shows that there is no one mechanism that is both private, accurate and cheap and the key remaining open problem is composing mechanisms under the small budgets of IoT hardware. Future work priorities involve adaptive privacy budgeting, communication and energy efficient secure aggregation, robustness to poisoning and standardised and reproducible evaluation.