Seminar Presentations
Präsentationen und Vorträge aus unseren Seminaren.
Alle Seminare
Federated and Differentially Private Hyperparameter Search
23. April 2026
Vortragende:r
Johannes Liebenow
"Selecting the correct hyperparameters can significantly impact the performance of a federated learning model. However, since these choices are evaluated using sensitive client data, adjusting them can result in the leakage of private information. Our algorithm DP-Hype successfully identifies optimal hyperparameters while safeguarding the privacy of participating clients based on Differential Privacy."
Hardware-accelerated NTT: New Perspective
17. Juni 2025
Saleh Mulhem
Towards a Unification of Reconstruction Attacks on Syntactic Privacy Models
05. Juni 2025
Niklas Zapatka
A Tale of Fully-Homomorphic Encryption and its Applications in Healthcare
21. Juni 2024
Saleh Mulhem
Optimizing for Statistical Independence using a KNN Density Estimator
23. Mai 2024
Kathleen Anderson
Feature extraction as a primer for privacy-preserving medical data analysis: Example approaches for facial video data
11. April 2024
Nele Brügge
DP Helmet: Distributed Non-Interactive Privacy-Preserving Learning of Convex Optimization problems
14. März 2024
Moritz Kirschte
The Principles of the GDPR (Die Prinzipien der DS-GVO)
01. Februar 2024
Herr Bruegger und Herr Zwingelberg vom ULD (Unabhängiges Landeszentrum für Datenschutz)
Uzl-Psychology in AnoMed
TBD 2024
Jonas Obleser
Grundlagen Clinical studies: current concepts, challenges and opportunities
09. November 2023
Jens Fiehler, CEO of eppdata
S-GBDT: Differentially Private Training of Gradient Boosting Decision Trees
26. Oktober 2023
Thorsten Peinemann und Moritz Kirschte
Federated and Differentially Private Hyperparameter Search
23. April 2026
Beschreibung
Selecting the correct hyperparameters can significantly impact the performance of a federated learning model. However, since these choices are evaluated using sensitive client data, adjusting them can result in the leakage of private information. Our algorithm DP-Hype successfully identifies optimal hyperparameters while safeguarding the privacy of participating clients based on Differential Privacy.
Privacy-Preserving Data Synthesis
09. April 2026
DRAMatic Speedup: Accelerating HE Operations on a Processing-in-Memory System
12. März 2026
Beschreibung
We present DRAMatic, which implements foundational HE operations on UPMEM's PIM system with arithmetic optimizations for large parameters. Evaluated against Microsoft SEAL, DRAMatic significantly closes the runtime and energy efficiency gaps, though it remains constrained by multiplication performance and data transfer overhead.
Privacy-Friendliness of Feature Extractors: Empirical Insights, Metrics and Correlations
29. Januar 2026
Beschreibung
Feature extraction is widely used to improve utility in differentially private (DP) classification, often under the assumption that pretrained or foundation models inherently provide privacy-friendly representations. In this presentation, we challenge this assumption through an empirical study of feature extraction methods across multiple image classification datasets. Our aim is to analyse how dataset metrics relate to downstream DP performance and what insights they offer into dataset characteristics. Finally, we apply these metrics to more challenging facial classification datasets. First findings suggest that strong features are not necessarily privacy-friendly features, emphasising the importance of a more systematic evaluation of representation learning in privacy-sensitive settings.
Answering Many Database Queries With High Privacy and Precision
18. Dezember 2025
Beschreibung
Determining the influence of adversaries with limited background knowledge in differential privacy (statistical privacy) is challenging, especially when the distribution of the released answers becomes complex. This difficulty increases for composed queries, since each query is conditioned on the previous answers. We provide a simple and general method to circumvent this issue and enable straightforward calculation of statistical privacy. Furthermore, we introduce a framework to handle different kinds of sampling in this context.
Hardware-accelerated NTT: New Perspective
17. Juni 2025
Beschreibung
Lattice-based Homomorphic Encryption schemes have become the key methodologies for today’s and the future’s secure world. It comes at the cost of a vastly increased computational load due to the multiplication of wide-integer coefficient polynomials. NIST recommends Number Theoretic Transform (NTT) as an efficient remedy. Nevertheless, NTT strongly requires acceleration for large numbers of coefficients. We explore a new hardware concept to build NTT accelerators and determine an optimal hardware architecture configuration across various problem sizes.
Towards a Unification of Reconstruction Attacks on Syntactic Privacy Models
05. Juni 2025
Beschreibung
Syntactic privacy models such as k-anonymity attempt to sanitize data sets such that privacy breaches are prevented. While many attacks are described in the literature, practitioners hesitate to adopt alternatives such as differential privacy, arguing that these attacks rely on exceptional cases which experts avoid in practice. We propose a formal model of syntactic privacy and derive a reconstruction attack which subsumes many existing attacks, thus simplifying the discussion and formal analysis of these attacks.
Towards Learning Differentially Private Probabilistic Relational Models II
08. Mai 2025
Beschreibung
Probabilistic relational models (PRMs) provide a well-established formalism to combine first-order logic and probabilistic models. By reasoning over groups of indistinguishable objects (PRMs abstract from individuals) and thus are a promising formalism to generate synthetic relational data that can be made publicly available without violating the privacy of individuals. Building on the previous talk, we take a closer look at how to efficiently learn a PRM from a given propositional probabilistic model.
Towards Learning Differentially Private Probabilistic Relational Models I
24. April 2025
Beschreibung
Probabilistic relational models (PRMs) provide a well-established formalism to combine first-order logic and probabilistic models. By reasoning over groups of indistinguishable objects (PRMs abstract from individuals) and thus are a promising formalism to generate synthetic relational data that can be made publicly available without violating the privacy of individuals. We investigate how a PRM can be learned from a given propositional probabilistic model and outline the use case of relational data synthesis using PRMs.
A Tale of Fully-Homomorphic Encryption and its Applications in Healthcare
21. Juni 2024
Optimizing for Statistical Independence using a KNN Density Estimator
23. Mai 2024
Feature extraction as a primer for privacy-preserving medical data analysis: Example approaches for facial video data
11. April 2024
DP Helmet: Distributed Non-Interactive Privacy-Preserving Learning of Convex Optimization problems
14. März 2024
The Principles of the GDPR (Die Prinzipien der DS-GVO)
01. Februar 2024
Uzl-Psychology in AnoMed
TBD 2024
Grundlagen FPGAs und FPGA Designs
23. November 2023
Grundlagen Clinical studies: current concepts, challenges and opportunities
09. November 2023
S-GBDT: Differentially Private Training of Gradient Boosting Decision Trees
26. Oktober 2023
Gaussian Processes and Differential Privacy
06. Juli 2023

