Overview:

The MLens project has explored a series of novel learning methods—so-called Differentially Private Learning Algorithms—which, given a sufficiently large dataset, prevent any attack on the training data. Within MLens, we investigated, implemented, and theoretically analyzed scalable distributed learning algorithms, time-series learning approaches, and novel foundational data-science methods such as clustering.

Through our work in MLens and the resulting advances in privacy-preserving distributed and machine learning, we have made a significant contribution to strengthening digital self-determination.

Results:

Distributed DP Helmet [1] introduced the first scalable machine-learning approach in which a single message per participant is sufficient to collaboratively train a model.

With DPM [2], we developed a new clustering approach that enables the discovery of robust statistical structures in sensitive data. We conducted an in-depth theoretical analysis of this strategy, identifying its strengths and limitations, thereby establishing mathematical guarantees for its applicability [3].

For the synthesis of time-series data, we proposed the DP Inductive Miner [4], a novel method that learns temporal dependencies through process-tree representations.

In collaboration with our partners UKSH and Zühlke, we additionally conducted complementary case studies on mortality prediction and distributed learning.


[1] Moritz Kirschte, Sebastian Meiser, Saman Ardalan, Esfandiar Mohammadi. Distributed DP-Helmet: Scalable Differentially Private Non-interactive Averaging of Single Layers. arXiv preprint. https://arxiv.org/abs/2211.02003

[2] Johannes Liebenow, Yara Schütt, Tanya Braun, Marcel Gehrke, Florian Thaeter, Esfandiar Mohammadi. DPM: Clustering Sensitive Data through Separation. In: ACM CCS '24, 2024. https://arxiv.org/abs/2307.02969

[3] Yara Schütt, Esfandiar Mohammadi. Understanding the Theoretical Guarantees of DPM. arXiv preprint. https://arxiv.org/abs/2506.18685

[4] Max Schulze, Yorck Zisgen, Moritz Kirschte, Esfandiar Mohammadi, Agnes Koschmider. Differentially Private Inductive Miner. In: 2024 6th International Conference on Process Mining (ICPM), 2024. https://arxiv.org/abs/2407.04595