Robust Machine Learning

What it is about

You learn the basics and current methods of robust machine learning (ML). The module covers types of attacks and threat models, practical attack methods, as well as empirical and certifiable defenses. In the end you will be able to assess methods for securing ML systems and select appropriate algorithms for concrete problems.

What you will be able to do

• Understanding adversarial basics, attack types and threat models
• Knowledge of attack methods (gradient-based, score-based, black-box)
• Familiarity with empirical defense methods, e.g. adversarial training
• Understanding robust methods for discrete domains (Graphs, LLMs)
• Knowledge of certifiable defenses (e.g. randomized smoothing, IBP)
• Ability to evaluate algorithmic pros and cons and select methods

What the module consists of

• VorlesungPresentation and discussion of theoretical concepts
• ÜbungenApplication of concepts for practice and illustration of properties
• Assignments / ProgrammierprojekteIndependent deepening and practical implementation of the topics

Teaching method

• Vorlesung mit Präsentationsfolienfor conveying and discussing theoretical concepts
• Übungenfor practicing the concepts and illustrating properties
• Assignments (inkl. Programmierprojekte)for individual self-study and practical application