Website
Curriculum

Modul CS5075-KP06

Trustworthy AI (TrustAI)

Duration:


1 Semester
Turnus of offer:


each summer semester
Credit points:


6
Course of studies, specific field and terms:
  • Master Computer Science 2019 (optional subject), Elective, Arbitrary semester
  • Master Medical Informatics 2019 (optional subject), ehealth / infomatics, 1st or 2nd semester
  • Master IT-Security 2019 (optional subject), IT Security and Privacy, 1st, 2nd, or 3rd semester
Classes and lectures:
  • CS5075-Ü: Trustworthy AI (exercise, 1 SWS)
  • CS5075-V: Trustworthy AI (lecture, 3 SWS)
Workload:
  • 20 Hours exam preparation
  • 100 Hours private studies
  • 60 Hours in-classroom work
Contents of teaching:
  • Guiding principles of Trustworthy AI: lawful, ethical and robust AI
  • Trustworthy Computing Basics: Security, Privacy, Dependability, Safety, Transparency, Explainability, Traceability, Accountability
  • De-anonymization methods using machine learning models
  • Mathematical notions for privacy-preserving machine learning methods
  • Privacy-preserving machine learning methods
  • Analyse maschinell gelernter Modellen (Robustness Check, Explainability)
  • Verifikation maschinell gelernter Modellen ((Statistical Testing), Model Checking)
  • Black-Box methods for extracting machine learning models (for economical reasons, for analysis, and for verification)
  • Attacks for manipulating machine learning models (adversarial examples, backdoors)
  • Hardening of machine learning methods against manipulation methods
  • • Robust machine learning methods against manipulation attacks Secure and privacy-preserving distributed learning methods (Privacy-Preserving Federated Learning)
Qualification-goals/Competencies:
  • All current techniques taught in the module and described above can be named and defined by the students and their functional proofs can be explained on the basis of applications.
  • The formal foundations from the course can be precisely explained
  • Students are able to identify advantages and disadvantages of planning and acting approaches
  • Understanding about potential vulnerabilities of machine learning methods w.r.t. privacy-violations and manipulation possibilities
  • Understanding of hardening methods compared to deanonymization and manipulation methods
  • Students can analyze complex security requirements
Grading through:
  • Oral examination
Is requisite for:
Responsible for this module:
Teachers:
Literature:
  • C. Dwork, A. Roth: The Algorithmic Foundations of Differential Privacy - Now Publishers Inc, 2014
  • Andrej Bogdanov: Lecture notes by Andrej Bogdanov from Chinese University of Hong Kong
  • : Current conference and journal articles on the topics of the event will be announced at the beginning of the event in the case of the seminar and at the discussion of the topic in the case of the lecture.
Language:
  • offered only in English
Notes:

Admission requirements for taking the module:
- None

Admission requirements for participation in module examination(s):
- Successful completion of exercises and project tasks as specified at the beginning of the semester.

Module Exam(s):
- CS5075-L1: Trustworthy AI, oral examination, 100% of module grade.

According to the decision of the examination board of computer science from 19.1.2022 this module can be chosen for Master SGO from WS 2019 in the area 5. elective.

Letzte Änderung:
1.2.2022