Computer Science
In intelligent and autonomous systems, users must often share private data with external platforms to enable advanced functionalities. A key challenge in this process arises when the data to be protected is not a single, precise value, but is instead inherently uncertain or represents a range of possible values. This thesis addresses the privacy of such uncertain data, which is formally represented and processed using set-based methods. We review existing privacy-preserving techniques and introduce new mechanisms to safeguard this data during processing on untrusted platforms, exploring both cryptographic and non-cryptographic approaches.
Specifically, the thesis presents novel privacy-preserving mechanisms for set-based data processing, categorized by the data type within the sets.
For sets of real-valued data, we introduce a differential privacy mechanism for set-based estimation in linear and non-linear dynamical systems. This approach protects the sensitive information contained within sets, such as zonotopes that model system uncertainties, while minimizing the loss of utility for the estimation process.
For sets of discrete, binary data, we propose a mechanism that uses Fast Fully Homomorphic Encryption to ensure privacy. This method allows for the secure processing of sets represented as logical zonotopes in untrusted environments, maintaining computational practicality.
The evaluations demonstrate the effectiveness and practicality of the proposed mechanisms across various applications in autonomous and intelligent systems.
Underwater robots need reliable perception for navigation, mapping, diver interaction, and manipulation, yet vision is degraded by wavelength-dependent attenuation, scattering, and variable water optics. These effects reduce contrast, distort color, and destabilize visual cues, so perception must be tailored to underwater image formation and field reliability constraints.
This thesis develops multimodal, 3D-aware perception for adverse marine and deep-sea conditions, based on experiments and integration within the EU projects MORPH, CADDY, and DexROV. By combining complementary sensors (2D imagery, stereo 3D structure, inertial and acoustic cues) with learning pipelines, the approaches compensate for individual sensor weaknesses.
First, it enriches 2D perception with 3D context and underwater-specific enhancement. Contributions include terrain-complexity estimation from texture metrics and stereo geometry to adapt AUV speed during surveys, plus color restoration/image enhancement to improve detection and pose estimation. For human-robot interaction, it introduces diver detection and pose estimation that merge stereo point-cloud descriptors with recurrent neural networks to handle low-contrast imagery.
Second, it presents end-to-end systems, including the CADDY underwater stereo-vision dataset for gesture-based communication and a gesture-recognition pipeline that blends classical learning, deep detectors, and a grammar-guided human-in-the-loop design for safer diver and AUV communication.
Finally, for deep-sea intervention, it proposes a simulation-in-the-loop validation to reduce sim-to-real gaps and an adaptive localization framework fusing dense 3D reconstruction, planar geometry, image-quality cues, and visual odometry to maintain accurate navigation in low visibility. The methods are validated on real data and integrated into autonomous demonstrators for safety-critical missions during field trials.
Mapping challenging confined underwater environments pushes the boundaries of what is possible for state-of-the-art robotics. Current state-of-the-art high-performance equipment allows already for accurate mapping in such scenarios. However, these systems are often expensive. Affordable underwater robotic systems and sensors come with significantly reduced capabilities. Especially sonars are necessary for mapping unknown environments, due to cluttered water resulting in bad visibility for vision based sensors. Yet, affordable sonar sensors suffer from higher noise levels, reduced accuracy, and limited coverage. Consequently, developing methods to achieve reliable and accurate mapping of challenging environments with affordable hardware remains an open research question.
This thesis presents a Fourier-SOFT in 2D (FS2D) registration method for robust matching of high-noise 2D sonar scans. A Simultaneous Localization and Mapping (SLAM) framework designed to the unique challenges of affordable Mechanical Scanning Sonars (MSS) is presented, integrating this FS2D registration method.
In the context of the digitization of cultural heritage, the Bunker Valentin Memorial in Bremen is surveyed, and maps of its multiple basins are generated.
Additionally, this thesis contributes an open dataset with accurate ground truth for development and benchmarking 2D sonar navigation, mapping, and SLAM algorithms.
Overall, this thesis demonstrates that, when the unique characteristics of affordable hard ware are considered correctly, and the methods are designed accordingly, affordable underwater robots can effectively map and explore challenging, unknown environments.
The BlueAUV design, the FS2D registration method, SLAM framework for affordable hardware, and an openly available dataset provide a foundation for advancing robust mapping of challenging underwater environments within the research community.
The rapid growth of large-scale software systems has led to the adoption of automatic error reporting platforms collecting millions of crash reports from real users. Central to these reports is the stack trace — a record of function calls leading to failure — which serves as a crucial diagnostic resource. However, the sheer volume, diversity, and redundancy of reports create bottlenecks: developers are overwhelmed by duplicates and highly variable submissions from the same defect, impeding efficient issue resolution.
Existing deduplication and triage solutions in industry and academia mainly rely on string-matching, information retrieval, or graph-based heuristics. While efficient, string and IR methods often miss semantic and contextual nuances; graph-based models lose detail about individual reports, reducing accuracy. These limitations cause missed linkages between related errors and fragmentation of bug databases. The lack of scalable algorithms, real-world benchmarks, and advanced learning methods further restricts current tools.
This dissertation advances automation of error report processing via stack trace analysis. It introduces (1) hybrid similarity metrics extending traditional techniques, (2) deep learning models for robust similarity estimation, (3) aggregation strategies leveraging group-level information, (4) scalable solutions for industrial use, (5) the first models for automated developer assignment in stack trace–centered triage, and (6) methods for interpreting and highlighting the most informative stack frames. The research is validated on multiple proprietary and open datasets, including new benchmarks released as part of this work. Together, these contributions provide a unified, reproducible foundation for scalable, accurate, and actionable error report deduplication, grouping, assignment, and tooling in real-world software engineering.
This thesis investigates consistency challenges in distributed stream processing systems. Prior work on this topic has made significant progress, with many ideas being implemented in state-of-the-art Stream Processing Engines (SPEs). In this thesis, we focus on formal modeling to better characterize existing problems and explore potential improvements.
We introduce a formal model of delivery guarantees and show that deterministic SPEs can theoretically achieve lower latency than non-deterministic ones for exactly-once guarantee. This is supported by experimental results demonstrating that a novel deterministic implementation performs better than current alternatives.
The thesis also presents a formal model for substream management, identifying a lower bound on the additional network traffic required for detecting substream termination. A corresponding framework is implemented that meets this bound and demonstrates improved performance over existing approaches.
These results contribute formal foundations and practical techniques for improving the performance and predictability of distributed stream processing systems.
Tests and proofs are two main techniques in modern software verification. To test a program means running the program to check if its execution yields an expected outcome. To prove a program is to build a mathematical proof, showing the correctness of the program against its desired properties. In the traditional view, however, tests and proofs are considered as two incompatible techniques. They are often treated as warring siblings and mostly applied in isolation. The complementarity of tests and proofs — though not immediately apparent — has been relatively underexplored. Can their combination mitigate each other’s weaknesses while harnessing their respective strengths? This thesis tries to straddle the border between tests and proofs and suggest a concrete answer. It explores how the two approaches can collaborate with and mutually benefit one another. Three key contributions arise from this exploration. The first contribution is Proof2Test, a framework that transforms failed proofs into useful test cases, allowing programmers to use tests to debug failed proofs effectively. The second contribution consists of several proof-based test generation strategies, which use proofs to enhance both the efficiency and effectiveness of test generation. This thesis also extends SC with “loop unrolling”, considering not just zero or one but any number of iterations, up to a set limit. It also includes an empirical study to examine how much (if anything) testing strategies miss when they limit themselves to standard branch coverage and, conversely, how many more bugs we can find if we unroll loops. The last contribution of this thesis is an automatic program repair approach, Proof2Fix, which takes advantages of the proposed test generation methods to produce meaningful corrections to faults revealed by proofs.
Implications of Dataset Heterogeneity on Deep Learning Performance in Medical Image Segmentation
(2021)
This thesis is about medical image segmentation using deep learning, with a particular focus on the influence of the training data. The performance of deep learning algorithms is impacted by the training set quality and heterogeneity, here grouped into three categories: technical image quality, reference segmentations and study populations.
Different training strategies are compared for parotid gland segmentation in CT data. All yield robust segmentation results, also in the presence of artifacts, and outperform non-deep learning methods on public data. Typical errors coincide with regions of high inter-observer variability. Training on contours from clinical routine, and on curated contours yield similar accuracy and results.
Bladder, rectum and uterus are segmented in cone-beam CT data, that is noisier and less well-calibrated than CT data. Using CT data for augmenting the anatomical variability is proposed and found to improve the performance. Prior knowledge about the presence of typical artifacts is integrated into the data sampling. Curriculum learning seems promising to increase the robustness to the particular artifact.
The hippocampus is segmented in CT data. A CT-only approach for generating the training contours could facilitate the data collection, but it is found that MRI-based training contours yield significantly higher performance and lower uncertainty.
White matter hyperintensity lesions are an imaging biomarker linked to stroke and cognitive decline. It is shown that a single neural network can segment these lesions in heterogeneous MRI data with varying image quality and lesion loads, and for a wide range of training set compositions, generated by pooling and systematic sampling. A challenge is the co-occurrence of stroke lesions. An approach that uses stroke segmentations for guiding the sampling, but not for optimizing the training loss, is proposed and found to outperform other sampling approaches with respect to false positive detections.
Modeling, Estimating, and Visualizing Spatial and Temporal Uncertainty for Image-Guided Therapy
(2021)
We address different aspects of the problem of uncertainty in image-guided therapies. This is of great interest, because such uncertainty could have a substantial impact on public health. Significant uncertainties can occur in image segmentation and estimating deformations of anatomy. To model the uncertainty in the image segmentation aspect, we propose a novel stationary Gaussian process (GP)-based generative segmentation model. This segmentation model allows us to draw many possible image segmentations, which can be used for estimating and visualizing different aspects of the uncertainty in image-guided therapies. To enable drawing of many image segmentation samples efficiently, we propose a fast method for sampling from stationary GPs. To model the uncertainty in the aspect of estimating deformations of anatomy, we propose a novel spatiotemporal GP model for uncertainty-aware soft-tissue motion estimation using GP regression. The spatiotemporal GP formalism enables the estimation of anatomy displacements at any location, and for any time interval from measured motions that are sparse in space and time. The use of GP regression enables the quantification of uncertainty in the soft-tissue motion estimation result, which allows the amount of uncertainty in some aspects, e.g., registered planning medical images, of image-guided therapies or procedures governing the decisions of medical specialists to be conveyed. To convey the amount of uncertainty in the anatomy motion estimates, we propose novel motion uncertainty visualization methods. To showcase the use of the devised methods, we deploy them in the context of radiotherapy and image-guided soft-tissue intervention navigation. We expect that incorporating estimates of spatial and temporal uncertainty into the processing pipelines of image-guided therapy will eventually enable improved treatments, and thus, improved outcomes.
Medical imaging is nowadays capable of non-invasively displaying the human brain in manifold ways. Computer assisted analysis of such data has become a highly active area of research. Today, there exists a wealth of published methods addressing a broad spectrum of medical questions. Considering the maturity to which this field of research has grown by now raises expectations that most basic problems should actually be solved and even many of the more complex task should at least be manageable. However, looking at the role of image-based quantification in everyday clinical practice, draws a completely different picture. The de-facto standard approach in diagnostic radiology is a purely qualitative reading through human experts. Quantification, if used at all, is typically limited to simple measurements on single image-slices. Without debate, there exists a huge gap between what would be potentially possible as defined by the scientific state-of-the-art, and clinical reality.
This PhD thesis addresses this situation. It studies three general concepts aiming at objective image assessment: 1) quantification; 2) interactive segmentation; 3) interactive data-visualization. For each of these concepts, an exemplary application is chosen, and a novel method is proposed with a focus on fulfilling requirements that, if not met, would prevent integration into clinical workflows. First, I present a method for robust assessment of upper-spinal cord atrophy, a parameter which has been successfully correlated to several clinical markers in the context of multiple sclerosis. Second, I present two novel interactive tools for segmenting individual gyri of the brain and regions-of-interest in DTI data. Finally, I present an example of how interactive 3D visualizations combined with efficient tools for exploration of image data can support the planning process of complex neurosurgical interventions.
Autonomous robotics in the real world is based on complex, interacting systems which demand for major integration efforts. In particular, the validation of such systems in order to reach production state raises the stakes in terms of parallelized testing and inter-workgroup interfacing. This thesis proposes to make use of a comprehensive simulation framework in order to increase the validation efficiency of such systems. Seamless integration into the testing pipeline allows for rapid development of components and validation under controlled conditions. This Continuous System Integration paradigm can be utilized to replace missing parts of the system, nevertheless ensuring realistic conditions, during the software development phase. Moreover, it is applicable during deployment in field trials as well where business logic can be replaced by simulated components with no overhead.
Additionally, with the foundations of a simulation framework being described and embedded into the context of system integration, this framework can further on be utilized for various knowledge expansion tasks, including benchmarking, optimization and autonomous reasoning needs. To open up these use cases, an abstract Simulation-Aided Knowledge Expansion concept is presented to run specified tasks in a loop within the simulation environment, making use of high-fidelity alignment with sensor data recorded in field trials. This concept allows for generating and expanding knowledge using the simulation in the sense that every iteration yields results which are supportive for use cases like benchmarking algorithms, optimizing parameters or increasing the basis of facts used in autonomous reasoning. For each of these use cases, an implementation of the concept is introduced and evaluated in detail.
