Federated Learning

Federated learning is the technique that lets an AI system improve using data that never leaves people's phones or a hospital's servers. Understanding it unlocks three things: (1) you will be able to read and judge privacy claims in AI products — does this really keep data local, or is it just marketing? (2) you will understand why some apps can improve their predictions without ever uploading your personal messages or health records; (3) you will be ready for the next steps in privacy engineering — differential privacy, which adds mathematical noise to protect individual data points, and secure aggregation, which lets multiple parties compute a shared result without seeing each other's inputs. These build directly on federated learning's core insight: share the update, not the data.

Think of a teacher who wants to improve a class assignment based on what students learned, but the students are not allowed to show their notebooks. Instead, each student does a small practice problem, then tells the teacher only one number: how much their answer would change if the lesson were adjusted. The teacher collects all these 'suggestions', combines them into a single adjustment, and updates the lesson for everyone. The notebooks never leave the students' desks. That is federated learning in essence: many devices (the students) each have private data (their notebooks). Instead of sending the data to a central server, each device trains a small copy of a shared model on its own data, then sends only the resulting update — a list of numbers showing how the model should change. The server averages these updates and improves the shared model. The raw data stays where it was born.

The key word is 'update'. An update is not the data itself — it is a change, a delta, like 'increase the weight for income by 0.2' or 'decrease the weight for debt by 0.1'. The server never sees the actual income or debt values; it only sees how the model should move. Over many rounds, the shared model becomes good at the task for everyone, even though no single person's private information was ever centralised.

A better analogy: a cooking class where the instructor wants to perfect a recipe, but each student has their own secret ingredient that they won't reveal. The instructor sends each student the current recipe. Each student cooks the dish at home, using their secret ingredient, and then reports back only one thing: 'If you increase the salt by a quarter teaspoon, the dish improves for me' or 'Decrease the pepper'. The instructor collects these suggestions from all students, averages them, and updates the recipe. After a few rounds, the recipe is better for everyone — but no student ever revealed their secret ingredient. This analogy holds well: the recipe is the model, the secret ingredient is the private data, the suggestion is the update, and the averaging is the server's job.

Where the analogy stops working: the students are giving *directions* (increase or decrease), but in real federated learning, the updates are precise numerical vectors — not just 'more salt' but 'change this specific weight by 0.03 and that one by −0.01'. Also, the instructor is not just averaging suggestions; they are running an optimisation algorithm to minimise a loss function. And crucially, in real federated learning, the data is not secret from the device itself — the device's own copy of the model can overfit to its local data if not handled carefully. The analogy also hides the problem of communication cost: in the cooking class, sending one number per student is cheap, but in federated learning, sending a full model update (which can be millions of numbers) is expensive — that is a central practical constraint.

Federated learning is a machine learning technique in which a shared model is trained across many participating devices or institutions that each keep their own local data, and only exchange model updates (not raw data) with a central coordinating server.

You are learning this as a foundation in privacy engineering. Federated learning is one of the four key techniques you'll study here — alongside data minimisation (not collecting data in the first place), anonymisation (removing identifiers), and differential privacy (adding noise for formal guarantees). Federated learning is different from all of those because it changes *where* training happens, not *what* data is collected. It does not guarantee privacy by itself — as your library notes, updates can still leak information, which is why differential privacy is often combined with it. You haven't yet studied neural networks, but federated learning sits on top of standard gradient-descent training; you'll meet that later. For now, think of the 'model' as a black box that makes predictions, and the 'update' as a set of numbers that improve it. This concept builds on the idea of distributed systems — many computers working together — but you don't need any prior knowledge beyond that.