004 Datenverarbeitung; Informatik
Refine
Year of publication
Document Type
- Doctoral Thesis (124)
- Article (25)
- Conference Proceeding (7)
- Other (3)
- Master's Thesis (1)
- Preprint (1)
Has Fulltext
- yes (161)
Is part of the Bibliography
- no (161)
Keywords
- Computersicherheit (7)
- Graphenzeichnen (7)
- Software Engineering (6)
- Graph (5)
- Multimedia (5)
- Programmanalyse (5)
- machine learning (5)
- Information Retrieval (4)
- Modellierung (4)
- Qualitätssicherung (4)
Institute
- Fakultät für Informatik und Mathematik (88)
- Mitarbeiter Lehrstuhl/Einrichtung der Fakultät für Informatik und Mathematik (50)
- Sonstiger Autor der Fakultät für Informatik und Mathematik (8)
- Wirtschaftswissenschaftliche Fakultät (8)
- Sozial- und Bildungswissenschaftliche Fakultät (3)
- Geistes- und Kulturwissenschaftliche Fakultät (2)
- Mitarbeiter Lehrstuhl/Einrichtung der Wirtschaftswissenschaftlichen Fakultät (2)
- Institut für IT-Sicherheit und Sicherheitsrecht (ISL) (1)
- Juristische Fakultät (1)
Smart Grids are electrical grids that require a decentralised way of controlling electric power conditioning and thereby control the production and distribution of energy. Yet, the integration of Distributed Renewable Energy Sources (DRESs) in the Smart Grid introduces new challenges with regards to electrical grid balancing and storing of electrical energy, as well as additional monetary costs. Furthermore, the future smart grid also has to take over the provision of Ancillary Services (ASs). In this paper, a distributed ICT infrastructure to solve such challenges, specifically related to ASs in future Smart Grids, is described. The proposed infrastructure is developed on the basis of the Smart Grid Architecture Model (SGAM) framework, which is defined by the European Commission in Smart Grid Mandate M/490. A testbed that provides a flexible, secure, and low-cost version of this architecture, illustrating the separation of systems and responsibilities, and supporting both emulated DRESs and real hardware has been developed. The resulting system supports the integration of a variety of DRESs with a secure two-way communication channel between the monitoring and controlling components. It assists in the analysis of various inter-operabilities and in the verification of eventual system designs. To validate the system design, the mapping of the proposed architecture to the testbed is presented. Further work will help improve the architecture in two directions; first, by investigating specific-purpose use cases, instantiated using this more generic framework; and second, by investigating the effects a realistic number and variety of connected devices within different grid configurations has on the testbed infrastructure.
Archival collections contain an underutilized wealth of biodiversity data, encapsulated in government files and other historical documents. In 1845, the Bavarian government conducted a comprehensive national survey on the occurrence of 44 selected vertebrate species across the country. The detailed expert responses from 119 forestry offices, totalling 520 handwritten pages, have been preserved in the Bavarian State Archives. In this study, we digitized, annotated, geographically referenced, and published these historical records, making them widely available as data for research and conservation planning. Our dataset, openly accessible through the Global Biodiversity Information Facility (GBIF) and Zenodo, contains 5,467 species occurrence records from 1845. Besides the binary presence/absence data, we have also published the original textual survey responses, which contain rich qualitative information, such as species abundances, population trends, habitats, forest management practices, and human-nature relationships. This information can be further processed and interpreted to address a range of questions in historical and contemporary ecology.
Motivated by the planarization of 2-layered straight-line drawings, we consider the problem of modifying a graph such that the resulting graph has pathwidth at most 1. The problem Pathwidth-One Vertex Explosion (POVE) asks whether such a graph can be obtained using at most 𝑘 vertex explosions, where a vertex explosion replaces a vertex 𝑣 by deg(𝑣) degree-1 vertices, each incident to exactly one edge that was originally incident to 𝑣. For POVE, we give an FPT algorithm with running time 𝑂(4𝑘 ⋅ 𝑚) and an 𝑂(𝑘2) kernel, thereby improving over the 𝑂(𝑘6) kernel by Ahmed et al. [2] in a more general setting. Similarly, a vertex split replaces a vertex 𝑣 by two distinct vertices 𝑣1 and 𝑣2 and distributes the edges originally incident to 𝑣 arbitrarily to 𝑣1 and 𝑣2. Analogously to POVE, we define the problem variant Pathwidth-One Vertex Splitting
(POVS) that uses the split operation instead of vertex explosions. Here we obtain a linear kernel and an algorithm with running time 𝑂((6𝑘 + 12)𝑘 ⋅ 𝑚). This answers an open question by Ahmed et al. [2].
Finally, we consider the problem Π-VertexSplitting (Π-VS), which generalizes the problem POVS and asks whether a given graph can be turned into a graph of a specific graph class Π using at most 𝑘 vertex splits. For graph classes Π that can be dfined
in monadic second-order graph logic (MSO2), we show that the problem Π-VS can be expressed as an MSO2 formula, resulting in an FPT algorithm for Π-VS parameterized by 𝑘 if Π additionally has bounded treewidth. We obtain the same result for the problem variant using vertex explosions.
[2]
R. Ahmed, S.G. Kobourov, M. Kryven, An FPT algorithm for bipartite vertex splitting, in: P. Angelini, R. von Hanxleden (Eds.), Graph Drawing and Network Visualization -30th International Symposium, GD 2022, in: Lecture Notes in Computer Science, vol.13764, Springer, 2022, pp.261--268.
With the increasing adoption of virtualization technologies across various industries, virtual machines (VMs) offer cost-effective solutions for obtaining computing power without the burden of initial investment or ongoing maintenance. However, the widespread use of VMs also increases the risk of malicious actors attempting to gain unauthorized access due to the possibility of accessing the VMs via standard internet protocols. Virtual Machine Introspection (VMI) and Forensic Memory Analysis (FMA) are two key cybersecurity methods for addressing these threats. While FMA leverages digital forensic techniques to extract and analyse information from system memory to explain security incidents, VMI typically works with live systems, analysing running processes to detect real-time threats. Both approaches face a significant challenge known as the ”semantic gap,” which arises from the need to infer high-level system information from low-level data such as physical memory and CPU registers.
This dissertation explores using machine learning to bridge the semantic gap in FMA and VMI applications. The research uses OpenSSH process heap dumps as a use-case to extract high-level structures, such as OpenSSH encryption keys, from raw process memory dumps. The study employs various techniques to isolate relevant memory sections, from basic memory chunking and entropy analysis to more advanced methods utilizing pointers and malloc headers. During this research study, we also identified the need for a foundation model in memory forensics. Foundation models are general purpose models trained on large amounts of data and users can later use these models to perform different tasks by finetuning the model. This research also addresses the challenge of detecting malware by analysing system-level API calls and employing custom feature engineering techniques. Given that the threat landscape is constantly evolving, we also investigate concept drift — a phenomenon where input data distribution changes affect predictive models’ performance. To mitigate the degradation in performance due to concept drift, we introduce a concept drift detection algorithm complemented by a custom sampling method that optimizes training data selection. This approach reduces the training dataset size by one-third, enhancing the efficiency of model training while maintaining high performance.
Generative pre-trained transformers (GPT) have recently demonstrated excellent performance in various natural language tasks. The development of ChatGPT and the recently released GPT-4 model has shown competence in solving complex and higher-order reasoning tasks without further training or fine-tuning. However, the applicability and strength of these models in classifying legal texts in the context of argument mining are yet to be realized and have not been tested thoroughly. In this study, we investigate the effectiveness of GPT-like models, specifically GPT-3.5 and GPT-4, for argument mining via prompting. We closely study the model's performance considering diverse prompt formulation and example selection in the prompt via semantic search using state-of-the-art embedding models from OpenAI and sentence transformers. We primarily concentrate on the argument component classification task on the legal corpus from the European Court of Human Rights. To address these models' inherent non-deterministic nature and make our result statistically sound, we conducted 5-fold cross-validation on the test set. Our experiments demonstrate, quite surprisingly, that relatively small domain-specific models outperform GPT 3.5 and GPT-4 in the F1-score for premise and conclusion classes, with 1.9% and 12% improvements, respectively. We hypothesize that the performance drop indirectly reflects the complexity of the structure in the dataset, which we verify through prompt and data analysis. Nevertheless, our results demonstrate a noteworthy variation in the performance of GPT models based on prompt formulation. We observe comparable performance between the two embedding models, with a slight improvement in the local model's ability for prompt selection. This suggests that local models are as semantically rich as the embeddings from the OpenAI model. Our results indicate that the structure of prompts significantly impacts the performance of GPT models and should be considered when designing them.
This study reports the Intraclass Correlation Coefficients of feedback ratings produced by OpenAI's GPT-4, a large language model (LLM), across various iterations, time frames, and stylistic variations. The model was used to rate responses to tasks related to macroeconomics in higher education (HE), based on their content and style. Statistical analysis was performed to determine the absolute agreement and consistency of ratings in all iterations, and the correlation between the ratings in terms of content and style. The findings revealed high interrater reliability, with ICC scores ranging from 0.94 to 0.99 for different time periods, indicating that GPT-4 is capable of producing consistent ratings. The prompt used in this study is also presented and explained.
Modern information retrieval (IR) systems, including web search engines and digital libraries, face challenges in simulating realistic user search behavior. Evolving interaction patterns and the integration of AI-powered interfaces make these challenges even harder. Traditional evaluation methods struggle to capture the dynamic nature of user interactions, particularly in complex search tasks and multi-stage information-seeking processes. User simulation offers a promising solution, providing a controlled environment for experimentation and allowing customization to model specific user behaviors and task contexts.
This research develops advanced techniques for user simulation in IR, creating more realistic and dynamic models than were previously possible. Key contributions include new methods for representing query reformulation, modeling how information needs change, and measuring the impact of different search environments on simulated user behavior. Specifically, this work introduces contextual Markov models, cognitive state models, and embedding space alignment techniques to accurately represent interactive search behavior.
Beyond model development, new evaluation methods and metrics are proposed for assessing the quality of simulated search sessions. These include statistical comparisons of session characteristics and classification-based approaches to distinguish between simulated and real user behavior. Additionally, this work leverages emerging technologies, such as large language models (LLMs) and retrieval-augmented generation, to improve the realism of user search behavior simulation.
The practical outcome of this research is a modular and extensible simulation framework. This framework incorporates advanced techniques like user type-specific Markov models, advanced query generation using LLMs, and conversational user models.
Election polls are frequently employed to reflect voter sentiment with respect to a particular election (or fixed-event). Despite their widespread use as forecasts and inputs for predictive algorithms, there is substantial uncertainty regarding their efficiency. This uncertainty is amplified by judgment in the form of pollsters applying unpublished weighting schemes to ensure the representativeness of the sampled voters for the underlying population. Efficient forecasting systems incorporate past information instantly, which renders a given fixed-event unpredictable based on past information. This results in all sequential adjustments of the fixed-event forecasts across adjacent time periods (or forecast revisions) being martingale differences. This paper illustrates the theoretical conditions related to weak efficiency of fixed-event forecasting systems based on traditional least squares loss and asymmetrically weighted least absolute deviations (or quantile) loss. Weak efficiency of poll-based multi-period forecasting systems for all German federal state elections since the year 2000 is investigated. The inefficiency of almost all considered forecasting systems is documented and alternative explanations for the findings are discussed.
An algorithm for boosting regression quantiles using asymmetric least absolute deviations, better known as pinball loss, is proposed. Existing approaches for boosting regression quantiles are essentially equal to least squares boosting of regression means with the single difference that their working residuals are based on pinball loss. All steps of our boosting algorithm are embedded in the well-established framework of quantile regression, and its main components – sequential base learning, fitting, and updating – are based on consistent scoring rules for regression quantiles. The Monte Carlo simulations performed indicate that the pinball boosting algorithm is competitive with existing approaches for boosting regression quantiles in terms of estimation accuracy and variable selection, and that its application to the study of regression quantiles of hedonic price functions allows the estimation of previously infeasible high-dimensional specifications.
Robust discrete optimization is a highly active field of research where a plenitude of combinations between decision criteria, uncertainty sets and underlying nominal problems are considered. Usually, a robust problem becomes harder to solve than its nominal counterpart, even if it remains in the same complexity class. For this reason, specialized solution algorithms have been developed. To further drive the development of stronger solution algorithms and to facilitate the comparison between methods, a set of benchmark instances is necessary but so far missing. In this paper we propose a further step towards this goal by proposing several instance generation procedures for combinations of min–max, min–max regret, two-stage and recoverable robustness with interval, discrete, budgeted or ellipsoidal uncertainty sets. Besides sampling methods that go beyond the simple uniform sampling method that is the de-facto standard to produce instances, also optimization models to construct hard instances are considered. Using a selection problem for the nominal ground problem, we are able to generate instances that are several orders of magnitudes harder to solve than uniformly sampled instances when solving them with a general mixed-integer programming solver. All instances and generator codes are made available online.