🚀 Try Zilliz Cloud, the fully managed Milvus, for free—experience 10x faster performance! Try Now>>

Milvus
Zilliz

How are updates synchronized in federated learning?

In federated learning, updates from distributed devices or servers (clients) are synchronized through a structured process of local training, aggregation, and global model updates. Clients train a shared machine learning model locally using their own data and send only the model updates (e.g., gradients or weight changes) to a central server. The server combines these updates to refine the global model, which is then redistributed to clients for further training. This cycle repeats until the model converges. Synchronization ensures that all clients contribute to a unified model without sharing raw data, balancing decentralization with coordinated learning.

The synchronization process typically follows a round-based communication protocol. For example, in each round, the server selects a subset of clients (like mobile devices or edge servers) to participate. These clients download the current global model, compute updates using local data, and transmit the updates back. The server aggregates these updates using methods such as Federated Averaging, which computes a weighted average of the clients’ model weights. The weights might be adjusted based on factors like the amount of data each client used, ensuring larger datasets have proportionally greater influence. This aggregation step is critical—it harmonizes divergent updates from clients operating on different data distributions, maintaining model consistency across the network.

Challenges in synchronization include handling stragglers (slow clients) and ensuring efficient communication. To address stragglers, servers often set a timeout or proceed once a minimum number of updates are received, avoiding bottlenecks. Techniques like gradient compression or sparsification reduce communication overhead by sending only significant updates. For instance, a server might prioritize clients with strong network connections in a given round to speed up aggregation. Privacy-focused methods, such as secure aggregation, further complicate synchronization by encrypting updates before transmission, requiring the server to decrypt only the combined result. These mechanisms ensure synchronization remains scalable and secure, even as the number of clients grows.

Like the article? Spread the word