Federated Gossip Protocol

Model updates propagate through the network using an epidemic or gossip-based communication model. Each client in a training round selects a small, random subset of peers and exchanges model parameters directly. This push-pull mechanism ensures that information eventually reaches all nodes with high probability, even in the presence of churn. The gossip averaging process converges to the true global average under mild connectivity assumptions, making it robust to dynamic network topologies.

Decentralized gossip protocols exhibit inherent resilience to Byzantine failures without requiring a trusted central coordinator. By leveraging robust aggregation rules—such as coordinate-wise median, trimmed mean, or Krum—at each peer, the system can tolerate a fraction of malicious nodes sending arbitrary or poisoned updates. This Byzantine resilience is critical for adversarial environments like wireless spectrum sensing networks where compromised devices may attempt to corrupt the collaborative learning process.

The protocol operates over an arbitrary peer-to-peer communication graph, which can be static, random, or time-varying. Convergence guarantees depend on the spectral properties of this graph—specifically the spectral gap of the mixing matrix. Common topologies include:

• Fully connected: Fastest convergence but highest communication cost
• Ring or grid: Minimal overhead with slower mixing
• Random geometric graphs: Realistic for wireless ad-hoc networks where physical proximity determines connectivity

Unlike synchronous federated averaging, gossip protocols naturally support asynchronous operation. Clients can initiate peer exchanges independently without waiting for a global synchronization barrier. This eliminates the straggler problem entirely—slow or intermittently connected devices do not delay the overall training progress. The system continues to converge as long as each node participates in gossip exchanges at a minimum rate, making it ideal for heterogeneous edge hardware in federated wireless learning deployments.

Gossip protocols can be combined with local differential privacy to provide formal privacy guarantees without a trusted aggregator. Each client adds calibrated noise—typically from a Gaussian or Laplacian distribution—to its model update before gossiping. The privacy-utility trade-off is governed by the privacy budget (ε, δ), and the decentralized nature of gossip averaging amplifies privacy through privacy amplification by iteration, where repeated randomized exchanges strengthen the overall guarantee beyond what single-round mechanisms provide.

In wireless federated learning, gossip protocols map naturally onto over-the-air computation (AirComp) techniques. Multiple clients can simultaneously transmit their model updates over a shared multiple-access channel, exploiting the waveform superposition property of the physical layer. The receiver obtains a noisy sum of the transmitted updates directly, without decoding individual signals. This analog aggregation dramatically reduces communication latency and bandwidth requirements, enabling efficient gossip rounds in dense wireless sensor networks.