Mixed-Timescale Differential Coding for Downlink Model Broadcast in Wireless Federated Learning

In federated learning (FL), the parameter server must broadcast the global model to devices in every iteration. While differential coding (DiC) can reduce communication overhead by transmitting only the changes between consecutive models, it is highly fragile: if a device misses a single update, it cannot reconstruct subsequent models, forcing it to train on outdated information. This paper addresses how to maintain communication efficiency while ensuring robustness against unreliable wireless downlink channels.

The authors introduce Mixed-Timescale Differential Coding (MTDC), a hierarchical compression strategy that operates at three distinct levels of quantization and temporal frequency:

1. Full-Model Broadcast: A high-resolution, infrequent transmission that serves as the primary synchronization point.
2. First-Level Differential Update: An intermediate-frequency update that uses an older, more reliably received model as its reference.
3. Second-Level Differential Update: A high-frequency, low-bitrate update that uses the most recent model as its reference.

By choosing different reference models for these updates, the system creates a fallback mechanism. If a device misses a high-frequency update, it can still reconstruct the global model using the first-level update or the full-model broadcast. The authors also propose an age-aware extension (A-MTDC) that dynamically adjusts the broadcast type based on the staleness of the model at the device and the observed failure statistics.

Standard FL communication strategies often assume reliable links or suffer from significant performance degradation when packets are lost. By enabling robust model reconstruction, MTDC allows for aggressive data compression without the catastrophic error propagation typical of standard differential coding. This is critical for practical wireless FL deployments where channel conditions are time-varying and unpredictable, ensuring that devices remain synchronized with the global model even under adverse conditions.