I am thrilled to share the culmination of my Master’s studies in International Software Systems Science at the University of Bamberg, where I achieved a 1.0 (Excellent) grade for my thesis: Evaluating Privacy-Utility Trade-offs in Differentially Private Mobility Data Analysis.
The Privacy Crisis in Mobility Data
Mobility data—such as GPS trajectories and cellular traces—is essential for urban planning and public health. However, traditional anonymization techniques like pseudonymization and k-anonymity are fundamentally broken when applied to high-dimensional spatial data. My research demonstrates that true privacy requires mathematically rigorous guarantees, leading me to Differential Privacy (DP).
Methodology and Engineering
In my thesis, I engineered an evaluation framework over 6,320 experimental configurations across three diverse datasets (GeoLife, Madrid EDM, and Berlin TAPAS). The infrastructure utilized advanced geospatial data engineering with Uber’s H3 hexagonal tessellation and native NumPy vectorization to achieve O(1) loop-time execution.
I systematically evaluated the two foundational cryptographic mechanisms of DP:
- Laplace Mechanism (Pure ε-DP): Provides the strongest worst-case mathematical bounds by injecting noise scaled to L1 sensitivity.
- Gaussian Mechanism (Approximate (ε,δ)-DP): A cryptographically relaxed approach optimized for high-dimensional mobility vectors.
Core Discoveries
A major focus of the research was analyzing M-Clipping strategies for User-Level Privacy. I established that expanding clipping bounds in highly dense urban datasets yields mathematically negligible utility conservation while catastrophically increasing the risk of privacy bleeds.
Furthermore, I empirically exploded the myth that Gaussian mechanisms strictly dominate in spatial domains. At the per-query level across extensive statistical utility metrics (like Earth Mover’s Distance and Kendall’s Tau Correlation), Laplace and Gaussian mechanisms proved computationally equivalent with variance bounded to just 2-6%.
Below are detailed visualizations, data sets, and my defense materials, illustrating the PET (Privacy Enhancing Technologies) ecosystem and deep dive metrics from my evaluation.
Presentation: M-Clipping and Core Findings
Below is a presentation representing one of the core research questions (RQ) of the thesis: the M-clipping factor and its profound impact on privacy-utility trade-offs.