Trustworthy Distributed Learning
| Vortragende/r (Mitwirkende/r) | |
|---|---|
| Nummer | 0000003316 |
| Art | Vorlesung |
| Umfang | 5 SWS |
| Semester | Wintersemester 2026/27 |
| Unterrichtssprache | English |
| Stellung in Studienplänen | Siehe TUMonline |
| Termine | Siehe TUMonline |
- 12.11.2026 09:45-11:15 N2407, Seminarraum
- 17.11.2026 11:30-13:00 N2409, Seminarraum
- 19.11.2026 09:45-11:15 N2407, Seminarraum
- 24.11.2026 11:30-13:00 N2409, Seminarraum
- 26.11.2026 09:45-11:15 N2407, Seminarraum
- 01.12.2026 11:30-13:00 N2409, Seminarraum
- 08.12.2026 11:30-13:00 N2409, Seminarraum
- 10.12.2026 09:45-11:15 N2407, Seminarraum
- 15.12.2026 11:30-13:00 N2409, Seminarraum
- 17.12.2026 09:45-11:15 N2407, Seminarraum
- 22.12.2026 11:30-13:00 N2409, Seminarraum
- 07.01.2027 09:45-11:15 N2407, Seminarraum
- 12.01.2027 11:30-13:00 N2409, Seminarraum
- 14.01.2027 09:45-11:15 N2407, Seminarraum
- 19.01.2027 11:30-13:00 N2409, Seminarraum
- 21.01.2027 09:45-11:15 N2407, Seminarraum
- 26.01.2027 11:30-13:00 N2409, Seminarraum
- 28.01.2027 09:45-11:15 N2407, Seminarraum
- 02.02.2027 11:30-13:00 N2409, Seminarraum
- 04.02.2027 09:45-11:15 N2407, Seminarraum
Teilnahmekriterien
Beschreibung
This course explains the recently introduced federated learning setting and its multiple challenges. It will focus on data privacy, robustness against malicious client attacks, and the challenge of reducing the communication cost. The course will give an overview of the setting of decentralized learning and some of its challenges. A deep study of the recent machine learning algorithms and their fundamental properties, used to guarantee privacy, robustness, and communication efficiency, will be carried out.
More precisely, the course covers the following topics
- An introduction to distributed learning in its centralized, federated, and decentralized forms and the respective challenges arising.
- Basics of machine learning algorithms: gradient descent, stochastic gradient descent, linear and logistic regression, and basic neural network architecture
- Data privacy in distributed learning and the main tools to guarantee privacy, e.g., information-theoretic privacy through secret sharing and differential privacy.
- Schemes ensuring information-theoretic privacy in federated learning, their advantages and shortcomings.
- Combination of information-theoretic privacy and differential privacy in federated learning.
- Malicious client attacks in federated learning: known attacks and defense mechanisms.
- Designing robust federated learning schemes.
- Joint privacy and robustness in federated learning: challenges and schemes achieving them.
- Communication cost and model/gradient compression.
- A primer on decentralized learning and the challenges arising.
- An overview of schemes allowing robust and private decentralized learning and open problems for future research.
- A semester-long project requiring the students to implement state-of-the-art machine learning algorithms of their choice (fine-tuning LLMs, CNN, gradient descent ...) and making them privacy-preserving and or robust against attacks, and test the change in performance
More precisely, the course covers the following topics
- An introduction to distributed learning in its centralized, federated, and decentralized forms and the respective challenges arising.
- Basics of machine learning algorithms: gradient descent, stochastic gradient descent, linear and logistic regression, and basic neural network architecture
- Data privacy in distributed learning and the main tools to guarantee privacy, e.g., information-theoretic privacy through secret sharing and differential privacy.
- Schemes ensuring information-theoretic privacy in federated learning, their advantages and shortcomings.
- Combination of information-theoretic privacy and differential privacy in federated learning.
- Malicious client attacks in federated learning: known attacks and defense mechanisms.
- Designing robust federated learning schemes.
- Joint privacy and robustness in federated learning: challenges and schemes achieving them.
- Communication cost and model/gradient compression.
- A primer on decentralized learning and the challenges arising.
- An overview of schemes allowing robust and private decentralized learning and open problems for future research.
- A semester-long project requiring the students to implement state-of-the-art machine learning algorithms of their choice (fine-tuning LLMs, CNN, gradient descent ...) and making them privacy-preserving and or robust against attacks, and test the change in performance
Inhaltliche Voraussetzungen
Knowledge of optimization, e.g., convex optimization, and machine learning
Knowledge of programming languages, e.g., Python, and basic experience with deep learning libraries such as PyTorch for a solid project
Basic knowledge of mathematics, e.g., linear algebra, calculus, and probability theory
The course "Channel Coding" is recommended
Knowledge of programming languages, e.g., Python, and basic experience with deep learning libraries such as PyTorch for a solid project
Basic knowledge of mathematics, e.g., linear algebra, calculus, and probability theory
The course "Channel Coding" is recommended
Lehr- und Lernmethoden
Lectures: The lecturer uses slides and the blackboard (or an iPad) to explain the fundamental theoretical concepts. Lectures are designed to be interactive and to encourage the students to ask questions and initiate discussions with the lecturer and each other.
Tutorials: The lecturer or teaching assistants solve concrete problems to simplify the understanding of the theoretical content and help the students better understand the material.
Students’ presentations: Each group of three students nominates one of them to present a paper of their choice. The paper is recent research on topics discussed in class. (Groups of four students nominate two students, each presents a paper)
Poster session: A workshop-style session in which each group of three students nominates one of them to present a poster. The others would go around and ask questions about other posters. This session is designed to introduce the students to environments of international scientific events.
Throughout the semester, the students can ask for question-and-answer sessions in which the lecturer and teaching assistants come to answer the students' questions about the topics discussed in class. There will also be some on-demand lab sessions, where a tutor or the teaching assistants provide help to further the students' progress in the programming tasks.
During the lectures, we will use slides to present the main concepts and the blackboard (or a tablet) to write the explanation.
All slides and handwritten notes will be uploaded to Moodle.
Moodle will be used for communication between the lecturers and the students, and between the students.
During tutorials, an exercise sheet will be used.
We will use mentimeter.com for formative assessment and tweedback during the Q&A sessions.
Tutorials: The lecturer or teaching assistants solve concrete problems to simplify the understanding of the theoretical content and help the students better understand the material.
Students’ presentations: Each group of three students nominates one of them to present a paper of their choice. The paper is recent research on topics discussed in class. (Groups of four students nominate two students, each presents a paper)
Poster session: A workshop-style session in which each group of three students nominates one of them to present a poster. The others would go around and ask questions about other posters. This session is designed to introduce the students to environments of international scientific events.
Throughout the semester, the students can ask for question-and-answer sessions in which the lecturer and teaching assistants come to answer the students' questions about the topics discussed in class. There will also be some on-demand lab sessions, where a tutor or the teaching assistants provide help to further the students' progress in the programming tasks.
During the lectures, we will use slides to present the main concepts and the blackboard (or a tablet) to write the explanation.
All slides and handwritten notes will be uploaded to Moodle.
Moodle will be used for communication between the lecturers and the students, and between the students.
During tutorials, an exercise sheet will be used.
We will use mentimeter.com for formative assessment and tweedback during the Q&A sessions.