For most of the last decade, the ML playbook was simple: collect data, ship it to a central server, train, deploy. It worked, but it came with baggage \u2014 transfer costs, latency, and a compliance headache every time “personal data” and “cloud” showed up in the same sentence.<\/p>\n\n\n\n
Federated Learning (FL) flips the pipeline. Instead of moving data to the model, it moves the model to the data. Training happens on the device or institution where the data already lives \u2014 a phone, a hospital server, a car \u2014 and only the learned updates<\/em> travel back to a central coordinator. Raw data never leaves its source.<\/p>\n\n\n\n This isn’t a privacy add-on bolted onto existing ML infrastructure. It’s a different architecture, and it’s already running in production at Google, Apple, and a growing list of healthcare and finance systems.<\/p>\n\n\n\n FL is a distributed training paradigm where multiple clients \u2014 devices or organizations \u2014 collaboratively train a shared model without pooling their raw data. Each client trains a local copy of the model on its own dataset and sends back only:<\/p>\n\n\n\n What never leaves the client: photos, messages, medical records, transaction histories \u2014 any raw, identifiable data.<\/p>\n\n\n\n The core assumption behind the design: data has gravity, and it’s often cheaper, safer, and legally simpler to move computation to the data than the reverse.<\/p>\n\n\n\n Centralized ML:<\/strong> Device \u2192 raw data \u2192 central server \u2192 training \u2192 model. Every client’s data ends up in one place, which means one place to secure, one place that can leak, and one dataset transfer that has to clear legal review.<\/p>\n\n\n\n Federated Learning:<\/strong><\/p>\n\n\n\n The server never sees a single raw training example \u2014 only the aggregate of what everyone learned.<\/p>\n\n\n\n 1. Federated server (coordination)<\/strong> Initializes the base model, selects which clients participate in a given round, receives updates, runs the aggregation algorithm, and redistributes the improved model. Google’s Federated Learning stack, TensorFlow Federated, and Flower all implement this role.<\/p>\n\n\n\n 2. Federated clients (local training, edge computing)<\/strong> Each client downloads the current model, trains it on-device against local data, computes the resulting update, and sends that update \u2014 nothing else \u2014 back to the server. All computation happens at the edge; no raw data is transmitted.<\/p>\n\n\n\n 3. Update extraction and protection<\/strong> Before an update leaves the client, it’s typically wrapped in additional protection \u2014 secure aggregation or differential privacy \u2014 so the server can’t reverse-engineer individual contributions from what it receives. This is the layer that turns “we don’t send raw data” into an actual privacy guarantee rather than a design intention.<\/p>\n\n\n\n 4. Central aggregation<\/strong> The server combines incoming updates into a new global model. The standard algorithm here is FedAvg (Federated Averaging)<\/strong>, introduced by Google in 2017: each client’s local weights are combined into a weighted average, weighted by how much data that client trained on. The result is a new global model that has absorbed knowledge from every participant without any of them exposing their dataset.<\/p>\n\n\n\n No \u2014 not by itself. Sending gradients instead of raw data is a real improvement, not a silver bullet. Known attack vectors include:<\/p>\n\n\n\n That’s why production FL systems layer in additional defenses:<\/p>\n\n\n\nWhat Federated Learning actually is<\/h2>\n\n\n\n
\n
How it works, concretely<\/h2>\n\n\n\n
\n
The architecture, component by component<\/h2>\n\n\n\n
Centralized ML vs. Federated Learning<\/h2>\n\n\n\n
<\/th> Centralized ML<\/th> Federated Learning<\/th><\/tr><\/thead> Data location<\/td> Aggregated on one server<\/td> Stays on each device\/institution<\/td><\/tr> What’s transmitted<\/td> Raw data<\/td> Model weights \/ gradients<\/td><\/tr> Privacy exposure<\/td> High \u2014 single point of failure<\/td> Low \u2014 no raw data in transit<\/td><\/tr> Compliance surface<\/td> Large (data transfer, storage, retention)<\/td> Smaller (no raw data leaves origin)<\/td><\/tr> Training location<\/td> Central infrastructure<\/td> Edge devices \/ local servers<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n Frameworks worth knowing<\/h2>\n\n\n\n
\n
Does FL actually guarantee privacy?<\/h2>\n\n\n\n
\n