CS-E4740: Federated Learning

Catalog description

Federated learning trains machine learning models across many devices that never share their raw data. The course builds it from a single design principle: a federated system is a graph whose nodes hold local datasets, and the algorithms — FedGD, FedAvg, FedProx, clustered federated learning — follow from applying different optimization methods to generalized total variation minimization over that graph.

The course then takes the parts that actually stop practitioners: how many training rounds buy how much accuracy, how to find which devices should learn together when you cannot see their data, how many malicious participants a protocol survives, what differential privacy costs in utility, and when an exchanged model parameter is still personal data under the GDPR and the EU AI Act.

Objectives

On completion, students can

• model federated learning applications using graphs;
• use total variation minimization as a design principle for federated learning;
• design federated learning algorithms by applying optimization methods to total variation minimization;
• analyze the computational and statistical properties of federated learning algorithms.

Lectures (tentative)

Twelve recordings of about 25 minutes — one idea each. All lectures are recorded and public on YouTube; slides are in the course repository. The list below is the plan for spring 2027 and may still change.

1. From one device to many
2. The empirical graph
3. GTVMin
4. Bias, variance, and the price of pooling
5. Fixed points and contractions
6. Gradient descent is the bridge
7. Protocols: FedGD, FedAvg, FedProx, FedRelax
8. Rounds, error, and the model ladder
9. Federated clustering
10. Legal frameworks for FL
11. Robustness and poisoning
12. Privacy

Live sessions

Alongside the recordings, the spring 2027 edition runs live clinics on YouTube: the recordings stay canonical, the live sessions handle what a recording cannot — questions mid-derivation, live debugging, and Q&A. They are open to everyone and are recorded.

• Ask Me Anything about Federated Learning — Wednesday 18 November 2026, 16:15 Helsinki time
• Course Kickoff: What Federated Learning Averages — Monday 1 March 2027, 16:15 Helsinki time
• GTVMin from Scratch: The FL Design Principle — Wednesday 10 March 2027, 16:15 Helsinki time
• Fixed Points and Why FedGD Converges — Monday 22 March 2027, 16:15 Helsinki time
• FL Student Project: Weather Forecasting Kickoff — Monday 5 April 2027, 16:15 Helsinki time
• Federated Clustering, Live in a Notebook — Wednesday 7 April 2027, 16:15 Helsinki time
• GDPR and the AI Act for Federated Learning — Monday 12 April 2027, 16:15 Helsinki time
• Poisoning a Federated Model, Live — Wednesday 14 April 2027, 16:15 Helsinki time
• The Localization Attack, and What Epsilon Costs the App — Monday 26 April 2027, 16:15 Helsinki time

Textbook

A. Jung, Federated Learning: From Theory to Practice. Singapore: Springer Nature Singapore, 2026 (doi:10.1007/978-981-95-1009-2), with the preprint free on arXiv:2505.19183. The course also draws on Machine Learning: The Basics (Springer, 2022) for the recap in module 1.

Requirements and grading

5 ECTS, with a 10 ECTS extended version. The course is assessed on

• Weekly assignments. Python with numpy and scikit-learn, on real weather-station measurements from the Finnish Meteorological Institute.
• A project. A report and a recorded talk, with anonymous peer review.

Assignment deadlines are Sunday 21 March, 4 April, 18 April and 25 April 2027. The weighting is not fixed yet; it will be announced here and in MyCourses before the course starts. The free follow-along track carries no credits, graded project or peer review — those require Aalto enrolment.

Prerequisites

Empirical risk minimization and the basics of supervised learning; gradient descent; linear algebra; Python with numpy and scikit-learn. Graph theory is developed from scratch, and no prior knowledge of distributed systems is assumed.

Intended audience

Master's and doctoral students in computer science, data science and electrical engineering, and practitioners who need to train models on data that cannot be centralized — health, industrial, mobile and public-sector data.