Primary Question: How can DP be leveraged as a constraint to understand the hierarchical nature of concept learning in machine learning models? Can DP be used to improve model interpretability?

Secondary Questions:

• How can we effectively identify and analyze the learning of specific concepts under DP constraints?
• What metrics can be used to quantify concept acquisition?
• How does (the level of) privacy influence the order and granularity of learned concepts?

The thesis will involve a combination of theoretical analysis and empirical evaluation. Possible methodologies include:

• Implementation in Python.
• Qualitative analysis of learned concepts and identification of relevant subpopulations.
• Development of quantitative metrics to evaluate concept acquisition under varying DP constraints.

Required:

• Strong motivation
• Strong programming skills in Python
• Background in machine learning and deep learning

Preferred:

• Experience with privacy-preserving ML frameworks
• Familiarity with differential privacy fundamentals
• Knowledge of interpretable AI techniques
• Good mathematical foundations

• Access to computing resources and datasets for experimentation
• Support from experienced researchers in DP and interpretability
• Potential opportunities to publish findings in AI/ML conferences

Interested students should send the following to sarah.lockfisch@tum.de and jonas.kuntzer@tum.de:

• A brief motivation letter (1 page max)
• Your CV
• Academic transcripts
• Any relevant project samples or publications (if available)

For further questions, feel free to reach out to Sarah Lockfisch or Jonas Kuntzer.