Thesis Topics

Parameter spaces for system performance evaluations often explode quickly with increasing factors and levels. This is particularly the case when evaluating complex setups such as Cloud-based database management systems. Here, both the cloud environment and the DBMS provide vast numbers of configurable parameters. At the same time, each specific configuration and evaluation run is costly both in terms of time and financial costs. The aim of this thesis is to develop a model that takes into account the parameter space, the estimated costs associated with evaluation runs, and the statistical robustness of outcomes. First, the model needs to incorporate experimental optimizations such as Plackett–Burman designs to mitigate the dimensionality of full-factorial designs. Second, the model must define cost functions to estimate the impact of indivdual run configurations. Third, the model needs to incorporate the effect of samlping and repeated measurements on the confidence of the results. The model should then serve as a foundation for performance engineers to balance the completeness of the coverage of the parameter space, the costs of running the evaluations, and the statistical accuracy of the results. The resulting tool should thereby assist performance engineers in their work and orchestrate evaluation run configurations. This thesis project is provided in collaboration with benchANT GmbH.

Recording traces of queries sent to a database system allows to replay these traces later as an evaluation workload. Furthermore, traces can be altered and modified in order to enhance the trace-based benchmarking capabilities. The aim of this thesis is to give an overview of the state of the art of trace-based DBMS benchmarks and to explore which tools are currently available for popular open source DBMSs. Furthermore, the thesis should explore the feasibility of advanced capabilities such as altering, tailoring, and scaling traces as well as the use of machine learning techniques to augment recorded traces. This thesis project is provided in collaboration with benchANT GmbH.

Many database management systems have been heavily extended recently with vector stores and MCP support to enable agentic access. However, the performance characteristics of these new access types have not yet been studied to a sufficient extent. This includes query patterns, transactional behaviors, and data access patterns. The aim of this thesis is to address this gap, compare agentic workloads with traditional database benchmarks (e.g., TPC-C), and derive and appropriate benchmark to evaluate database systems for such agentic usage. This thesis project is provided in collaboration with benchANT GmbH.

Performance measurements, performance evaluations, and performance engineering play an important role when designing and implementing complex software systems and distributed architectures. If you are interested in this area and seek a potential topic, please contact me for further discussion and drafting.

Human psychology impacts users in their privacy behavior. This is relevant for understanding user behavior, but also when designing technical privacy solutions. If you are interested in this intersection of psychology and privacy and seek a potential topic, please contact me for further discussion and drafting.

Data-intensive systems manage and process large volumes of data. These systems come with inherent challenges in terms of scalability, parallelism, programming models, architectures etc. If you are interested in this area and seek a potential topic, please contact me for further discussion and drafting.

Large Language Models (LLMs) are transforming many aspects of society and IT systems. Large pre-trained models demonstrate impressive problem understanding and solving capabilities and generalize to domain-specific tasks through in-context learning without further task-specific fine-tuning. This has led to a fast-evolving area of research that focuses on using LLMs to automate complex workflows like software development, system configuration and security incident response. Despite significant progress in this field, the aspects of privacy, security and reliability of such LLM-based systems often remain underexplored. Especially in high-stakes domains like network access control and healthcare or when dealing with sensitive data, reliability, security and privacy of such systems is crucial. We offer various bachelor and master thesis projects in this area of research. Specific thesis topics can be individually discussed and designed with the supervisor.

This topic addresses the challenges of protecting sensitive information in open data. Even when datasets are anonymized, unintended information leakage can and does still occur. I am particularly interested in quantifying the extent of such leakage and developing privacy-preserving measures to mitigate it. This may also include exploring static code analysis techniques or machine learning approaches to identify potential privacy risks in datasets. If this sounds interesting to you, get in touch with me and we can then further discuss the specific focus and scope of the thesis.

This topic addresses the challenges of securing modern network environments through Zero Trust Network Access. Even when strong perimeter defenses are in place, the lack of true end-to-end network traffic separation still leaves systems vulnerable to lateral movement. I am interested in developing architectures that enforce strict network isolation and verification. This may also include exploring unified policy frameworks to seamlessly synchronize security rules across both the core network infrastructure and the client endpoints. If you wish to explore this topic further, I invite you to contact me and we can then further discuss the specific focus of the thesis.

System performance evaluations produce quantities of structured data sets from measurements. These data sets need to be further analyzed to derive actual insights. The aim of this thesis is to evaluate how humans can be assisted in such analyses by (smaller) large language models (LLMs) that can run locally on commodity hardware. However, as LLMs are primarily good in processing and generating text-based content, this may require additional tooling to handle structured data and statistical analyses. As part of the thesis, different approaches and concrete solutions should be explored and compared. This thesis project is provided in collaboration with benchANT GmbH.

Increasingly capable AI models cannot only be used to vibe code projects without any programming skills, but also to assist engineers when enhancing, refactoring, and securing existing code. The aim of this thesis is to evaluate how well such models can be utilized for improving the performance of existing applications by identifying and mitigating performance issues. The thesis should particularly focus on publicly available coding models that can be deployed locally (e.g., gpt-oss:20b, Qwen3.5-coder:35b, glm-4.7-flash, gemma4:31b). As part of thesis, appropriate use cases for benchmarking the models need to be identified from literature and preparated. This involves different performance issues in software code - both obvious and rather subtle and difficult to detect. The models should then be tested and evaluated against these cases and finally synthesized in a comparative summary.

Zero-Knowledge Proofs (ZKPs) are a highly useful building block for engineering privacy-enhancing systems, yet their practical application and implementation remain notoriously complex. Recently, several mature, well-maintained software libraries have emerged to abstract this complexity and streamline the integration of ZKPs. This thesis aims to conduct a multi-faceted comparative analysis of three prominent, generalized zkSNARK frameworks: arkworks (Rust), gnark (Go), and snarkjs/circom (JavaScript/WASM). The core practical component of this thesis will be the implementation of an anonymous credential scheme across all three frameworks, using a common baseline implementation as a reference. These implementations will provide the empirical basis on which to evaluate the frameworks across qualitative dimensions, such as developer experience and quality of documentation. In addition to this qualitative analysis, the resulting implementations should undergo a rigorous performance evaluation, which will be analyzed and discussed in consideration of existing public benchmarks.

Security mechanisms in Vehicle-to-Everything (V2X) environments, such as misbehavior detection systems, generate outputs that indicate potential malicious behavior but do not directly provide a unified and interpretable trust representation. This thesis investigates methods for quantifying such outputs into subjective logic opinions that can be used by trust assessment frameworks. The focus lies on a machine learning–based approach that learns the mapping from security mechanism outputs to belief, disbelief, and uncertainty values. The proposed method will be compared against existing quantification techniques to evaluate improvements in accuracy, robustness, and interpretability. The evaluation will be conducted using realistic V2X datasets and scenarios.

State-machine replication is a concept to achieve fault tolerance. Each replica has be executed in a deterministic way. In order to allow concurrency so called deterministic multithreading approaches were developed, one of them is called the UDS scheduler. The task of this work is to extend the UDS implementation in order to inject application knowledge into the scheduling decisions so that more concurrency can be achieved compared to no knowledge. The knowledge is injected by calling methods in the scheduler by the application. Multiple different of such methods are to be implemented (e.g., last lock, next lock) and integrated into the scheduling algorithm. Further especially for a Bachelor's thesis some evaluations should be applied in order to show the benefit of the injected knowledge. The work will need some acquaintance with the theoretical basics of the UDS scheduler. The implementation needs skills in Java.

This thesis will take a quantitative look into statistical practices in empirical research. By analyzing the code that is submitted with preprints on the common preprint server PsyArXiv, we will identify common statistical pipelines and practices. This will allow us to gain insights into common patters and potential pitfalls, which can then aid future research in developing better tools and guidelines for statistical analysis.

This thesis aims to investigate the reliability and trustworthiness of traffic sign classifiers when subjected to occlusions. Utilizing the German Traffic Sign Recognition Benchmark (GTSRB) dataset, this research will focus on annotating the dataset with various levels and types of occlusions to evaluate if the predictions are still trustworthy. The primary objective is to assess the performance degradation of the classifier under different occlusion scenarios and to develop strategies to enhance its robustness. This study is crucial for improving the safety and reliability of autonomous driving systems where traffic signs might be partially obscured.

The growing connectivity of cooperative intelligent transportation systems (C-ITS) raises critical security concerns, particularly regarding the trustworthiness of exchanged information. Misbehavior Detection Systems (MDS) play a key role in identifying malicious or faulty behavior in vehicular networks. This thesis evaluates both AI-based and non-AI-based MDS approaches, comparing their detection performance and computational efficiency under different attack scenarios. Using a structured benchmarking framework and representative vehicular datasets, the study analyzes different kinds of MBDs.

Modern vehicles rely on complex in-vehicle networks to support safety-critical and comfort functions. As these networks become more interconnected, they face increasing security risks from malicious attacks and faulty components. Intrusion Detection Systems (IDS) are essential for detecting abnormal behavior and protecting the integrity of in-vehicle communication. This thesis evaluates different IDS approaches for automotive networks, including rule-based and machine learning methods. Using representative datasets and realistic attack scenarios, the study compares detection accuracy, false positive rates, and computational efficiency

This thesis will investigate methods for detecting natural adversarial examples against ImageNet classifiers using classic computer vision techniques. Adversarial examples are inputs to machine learning models that are designed to cause the model to make a mistake. This project will utilize the Harder ImageNet Test Set (https://arxiv.org/abs/1907.07174) as an dataset for Natural Adversarial Examples. The primary objective is to explore and compare the effectiveness of traditional computer vision methods, such as histograms and SIFT (Scale-Invariant Feature Transform), in identifying these adversarial examples. The outcome of this research will enhance our understanding of model vulnerabilities and contribute to developing more robust machine learning systems.

This thesis will explore the effectiveness of different optimization strategies in the generation of adversarial patches. Adversarial patches are small, intentionally designed perturbations that can cause machine learning models, particularly in computer vision, to misclassify inputs. The primary objective of this research is to compare various optimization techniques, such as gradient-based methods, evolutionary algorithms, and reinforcement learning, to determine which methods are most effective and efficient in creating these patches. The outcome of this research could significantly enhance our understanding of model vulnerabilities and contribute to the development of more robust machine learning systems.

This thesis aims to investigate the reliability and trustworthiness of traffic sign classifiers when subjected to occlusions. Utilizing the German Traffic Sign Recognition Benchmark (GTSRB) dataset, this research will focus on annotating the dataset with various levels and types of occlusions to evaluate if the predictions are still trustworthy. The primary objective is to assess the performance degradation of the classifier under different occlusion scenarios and to develop strategies to enhance its robustness. This study is crucial for improving the safety and reliability of autonomous driving systems where traffic signs might be partially obscured.

This thesis will investigate methods for detecting natural adversarial examples against ImageNet classifiers using classic computer vision techniques. Adversarial examples are inputs to machine learning models that are designed to cause the model to make a mistake. This project will utilize the Harder ImageNet Test Set (https://arxiv.org/abs/1907.07174) as an dataset for Natural Adversarial Examples. The primary objective is to explore and compare the effectiveness of traditional computer vision methods, such as histograms and SIFT (Scale-Invariant Feature Transform), in identifying these adversarial examples. The outcome of this research will enhance our understanding of model vulnerabilities and contribute to developing more robust machine learning systems.

An autonomous vehicle is equipped with a variety of sensors that produce large quantites of data which the vehicle uses to run a lot of different safety-critical functions, such as Cooperative Adaptive Cruise Control or Park Assist. In this thesis, we focus on the trust between the vehicle computer and other in-vehicle components that it relies upon to provide non-compromised data as input to different safety-critical functions. The goal of the thesis is to build a tool that will automate building of in-vehicular trust models based on a system model of a vehicle. A system model of a simplified vehicle will first need to be created by using the System Modeling Language (SysML). This model will serve as an input to the automation tool that needs to output a trust model in a pre-defined form. The methodology for building such trust models will be provided.

This thesis will explore the effectiveness of different optimization strategies in the generation of adversarial patches. Adversarial patches are small, intentionally designed perturbations that can cause machine learning models, particularly in computer vision, to misclassify inputs. The primary objective of this research is to compare various optimization techniques, such as gradient-based methods, evolutionary algorithms, and reinforcement learning, to determine which methods are most effective and efficient in creating these patches. The outcome of this research could significantly enhance our understanding of model vulnerabilities and contribute to the development of more robust machine learning systems.