Fachbereich Wirtschaft, Informatik, Recht
Refine
Document Type
- Article (59)
- Master's Thesis (31)
- Article in a Periodical of the TH Wildau (30)
- Conference Paper (25)
- Part of a Book (11)
- Report (10)
- Conference Proceedings (9)
- Working Paper (9)
- Monograph (3)
- Other (3)
Year of publication
Institute
Keywords
- Künstliche Intelligenz (7)
- artificial intelligence (7)
- machine learning (7)
- AI (4)
- Bibliothek (4)
- Maschinelles Lernen (3)
- automatic speech recognition (3)
- citizen science (3)
- natural language processing (3)
- performance (3)
Internationale Studierende nutzen Hochschulwebseiten als erste Informationsquelle zu Studienangeboten, Bewerbungen, Zulassungen, Visa, Finanzierung, Wohnen und Studienorganisation. Diese Arbeit untersucht, welche Bedeutung der Nationale Kodex für das Ausländerstudium an Hochschulen und für internationale Studierende hat, wie ausgewählte kodexbezogene Informationen auf der Hochschulwebseite der Technischen Hochschule Wildau umgesetzt werden und wie internationale Studierende diese digitale Umsetzung wahrnehmen. Der Kodex wird dabei als Soft-Law-Instrument eingeordnet. Die explorative Untersuchung folgt einem Mixed-Methods-Ansatz. Der öffentlich zugängliche Stand der TH-Wildau-Webseite im zweiten Quartal 2026 wird beschrieben. Ergänzend werden ein Experteninterview mit dem International Office sowie eine Online-Befragung internationaler Studierender ausgewertet. Die Ergebnisse zeigen, dass der Kodex an der TH Wildau sowohl als Orientierungsrahmen als auch als Instrument der internen Qualitätssicherung dient. Für internationale Studierende ist er als eigenständiges Instrument kaum erkennbar. Er wird durch die Bereitstellung relevanter Informationen und Angebote auf der Webseite wahrnehmbar. Viele kodexbezogene Informationen sind vorhanden, jedoch nicht einheitlich, gebündelt, leicht zugänglich oder vollständig auf Englisch verfügbar. Die Befragung zeigt insgesamt eine positive Wahrnehmung der Webseite, verweist allerdings auch auf Herausforderungen bei Sprache, Struktur, Zugänglichkeit und Aktualität. Die Wirkung des Kodex entsteht erst durch eine verständliche, mehrsprachige, verlässliche und zugängliche digitale Umsetzung.
This research paper addresses the critical cybersecurity issues of healthcare IT systems. Through the use of artificial intelligence, specifically natural language processing (NLP) and advanced neural networks such as long short-term memory (LSTM), the study closely analyzes vulnerabilities documented in the National Vulnerability Database (NVD). The main objective is to identify patterns in the Common Vulnerabilities and Exposures (CVE) datasets to predict Common Vulnerability Scoring System (CVSS) scores with high accuracy, which is of utmost importance in the healthcare sector where IT failures can have catastrophic consequences.
The work aims not only to predict the potential impact on healthcare but also to prioritize the vulnerabilities according to their severity and their relevance to healthcare facilities. The vulnerability descriptions, vectors, and affected software configurations will be evaluated in detail. Through this analytical effort, the study will develop a differentiated framework for the early detection and professional management of IT security threats, aiming to strengthen the cybersecurity defenses of healthcare infrastructures. By contributing to the strategic anticipation and mitigation of risks, the research seeks to improve the resilience of critical healthcare systems and ensure the continuity and integrity of patient care.
XplainCVSS: Interpretable Machine Learning for Explainable CVSS Vulnerability Severity Prediction
(2026)
Automating CVSS scoring can reduce analyst workload, but practical deployment requires explanations that are not only interpretable, but also faithful to model decisions. We present XplainCVSS, a framework that predicts all CVSS v3.1 base metrics from NVD vulnerability descriptions and evaluates explanation faithfulness quantitatively. Using 173,202 CVEs with strictly temporal splits, we benchmark ten models from TF-IDF+XGBoost to domainspecific and multi-task transformers. Fine-tuned SecureBERT achieves the best performance (macro-F1 = 0.774; reconstructed base-score MAE = 0.934), while multi-task variants offer competitive consistency at 7–10x lower training cost. We validate TreeSHAP and gradient-based transformer attributions across all eight CVSS components, showing that highlighted features are decision-relevant. A temporal drift decomposition further shows that label-distribution shift (concept drift) is the main driver of degradation (delta-F1 = 0.079), indicating the need for periodic retraining. Riskaware evaluation yields a low 2.79% under-prediction rate for Critical vulnerabilities, supporting analyst-in-the-loop triage. Overall, XplainCVSS highlights a practical trade-off between transformer accuracy and the direct interpretability of keyword-based models.
Enterprise security teams must triage thousands of newly disclosed software vulnerabilities—catalogued as Common Vulnerabilities and Exposures (CVEs)—every day, yet the eventual remedy—vendor patch, workaround, mitigation, or no fix—is unknown at disclosure time, and actionable guidance for unpatched vulnerabilities is scattered across heterogeneous advisory sources at scale. We study this problem along three dimensions. First, we frame remedy prediction as four-class classification on 68,105 CVEs, showing that a concatenation multilayer perceptron (concat-MLP) fusing SecureBERT embeddings with structured metadata achieves F₁=0.793, outperforming text-only encoders by 9.3 pp; confidence thresholding yields 97.5 % accuracy at 94.3 % coverage. Second, we present a cross-source validation of remedy labels across Red Hat, Debian, and GitHub Security Advisories (GHSA): inter-source agreement is low (ϰ=0.007−0.369) and multi-source training cannot recover this gap, indicating systematic ontological disagreement. Third, we construct the CVE-Workaround Corpus (21,794 records from four sources) and benchmark five generation systems; surface metrics diverge sharply from semantic quality, and openweight models still lag behind frontier large language models (LLMs). Together, these contributions provide a predict-validate-act blueprint for vulnerability triage as a big-data service.
Network vulnerability assessment relies heavily on Nessus Attack Scripting Language scripts for automated security testing. Traditional approaches to generating these specialized security scripts face significant challenges in maintaining syntactic correctness and semantic relevance. This paper introduces a novel predictable template learning methodology that transforms non-deterministic Common Vulnerabilities and Exposures descriptions into deterministic Nessus Attack Scripting Language code generation through systematic placeholder preprocessing. We fine-tune DeepSeek Coder 6.7B using Low-Rank Adaptation on a carefully curated dataset of 99,143 Common Vulnerabilities and Exposures-Nessus Attack Scripting Language pairs, implementing template-based preprocessing to replace unpredictable elements with consistent placeholders. Our resulting fine-tuned model, NASLGenDSC, achieves 86 % training loss reduction with a final perplexity of 1.17 compared to zero-shot approaches. Comprehensive baseline evaluations against DeepSeek Coder, CodeLlama 7B, and StarCoder2 7B show our approach achieves 30-43% lower perplexity across 300 test evaluations, validating the effectiveness of template learning for security-critical code generation. The deterministic nature of our output enables practical deployment in production vulnerability management systems.
Role-Based Analysis of Public Administration in National AI Strategies Under Evolving AI Governance
(2027)
Artificial intelligence (AI) is rapidly reshaping public administration by influencing how governments design services, manage data, and exercise regulatory authority. National AI strategies have become essential tools for governments to articulate their ambitions and define their role in governing AI, yet the policy landscape has evolved significantly since their adoption. This paper examines how national AI strategies define the role of public administration in AI governance. Drawing from concepts in AI governance, digital government, and policy design, we propose a role-based analytical framework that conceptualizes public administration as AI adopter, regulator, and orchestrator of AI ecosystems. We apply the framework in a qualitative comparative policy analysis of national AI strategies in Germany, the Netherlands, and the United Kingdom. Using a deductive-inductive qualitative content analysis, we reconstruct country-specific role profiles and compare strategic emphases across the three cases. Our findings reveal that all three strategies foreground the adopter role, with extensive public-sector use cases and capacity-building measures, while systematic scaling beyond pilots is made explicit only in the UK’s “scan, pilot, scale” approach. Regulatory framings are more prominent and tightly coupled to EU-level developments in Germany and the Netherlands, whereas the UK emphasizes innovation and competitiveness with fewer detailed legal commitments. Orchestration arrangements likewise diverge, ranging from partnership-oriented ecosystem coordination in the Netherlands and the UK to more hierarchical steering in Germany. We argue that a role-based lens helps to uncover how early strategic choices position public administrations within emerging AI ecosystems and to identify misalignments between these framings and the evolving European AI governance landscape, thereby providing an empirically grounded lens to assess whether existing strategies remain fit for purpose.
Im Forschungsprojekt PhaceSpace wurde untersucht, wie sich Phantombilder mittels generativer, bilderzeugender KI (StyleGAN) erstellen lassen. Dazu wurde ein Prototyp entwickelt, der StyleGAN durch einen aufgabenspezifischen Mensch-KI-Dialog um einen iterativen Suchprozess im Gesichtsraum erweitert. In zwei Nutzerstudien wurden Benutzerschnittstelle und Interaktionsmöglichkeiten sowie der Such- und Auswahlprozess experimentell untersucht und weiterentwickelt. Im Ergebnis zeigte sich, dass die Nutzenden mit dem Prototyp geeignete Bilder von Zielpersonen erstellen konnten. Weiterentwicklungspotenzial besteht insbesondere in einer gezielten Suche und Auswahl nicht nur der Gesichter als Ganzes, sondern auch einzelner Gesichtspartien.
Der Beitrag zeigt Umfang und Inhalt der Anforderungen an Künstliche Intelligenz und der dazu errichteten Governancestruktur gemäß der KI-Verordnung der Europäischen Union. Danach werden grundlegende Mängel der Verordnung diskutiert, wie sie von Wissenschaft und Interessenträgern formuliert werden. Die Europäische Kommission hat die Mangelhaftigkeit des aktuellen Regelungsansatzes zum Teil selbst erkannt und zur Abhilfe einen Gesetzgebungsvorschlag zur Nachbesserung eingebracht (Digital-Omnibus-Verordnung zur KI). Ausgewählte zentrale Inhalte dieser Änderungsverordnung werden besprochen und gewürdigt. Im Ergebnis ist die grundlegende Kritik an der KI-Verordnung aufrechtzuerhalten.
Large Language Models (LLMs) bieten Potenzial zur Automatisierung wissensintensiver Dokumentenerstellung; die anwendungsspezifische Modellauswahl bleibt jedoch eine methodische Herausforderung, da generische Benchmarks die Eignung für qualitätskritische Dokumentationsprozesse nur unzureichend abbilden. Die vorliegende Arbeit entwickelt daher eine Systemarchitektur zur qualitätsgesicherten Dokumentenerstellung und evaluiert drei aktuelle Modelle – GPT-5, Claude Sonnet 4.5 und LLaMA 4 Maverick – anhand von sieben anwendungsspezifischen Kriterien, kombiniert aus automatisierten Systemmetriken, LLM-as-a-Judge und Human Evaluation. Die Ergebnisse offenbaren eine Kriterienhierarchie, die zwischen notwendigen Basisanforderungen – Faktentreue und zuverlässige Systemkomponenten-Integration – und differenzierenden Qualitätsmerkmalen unterscheidet und damit eine empirisch fundierte Grundlage für die Modellauswahl bereitstellt.
Purpose – This study aims to explore how Generation Z employees in South Korea perceive and experience quiet quitting (QQ) and to examine the work-context factors associated with this phenomenon.
Aim(s) – The aim of this study is to clarify the impact of the phenomenon of "Quiet Quitting" among the Generation Z employees in South Korea.
Design/methodology/approach – A qualitative research design was adopted. Data were collected through semi-structured, in-depth interviews with eight South Korean Generation Z employees and analysed using a seven-step grounded analysis approach.
Findings – The findings suggest that, although generational differences remain a source of workplace tension, long working hours and the pursuit of work–life balance are among the most influential factors shaping participants’ attitudes toward QQ. The study further highlights the relevance of cultural factors, including hierarchical workplace structures and attitudes toward hustle culture, in understanding how QQ is experienced by South Korean Gen Z employees. Participants generally viewed QQ not as disengagement from work, but as a means of protecting themselves from excessive workloads, burnout, and perceived imbalances between effort and reward.
Limitations of the study – The study is limited by its small sample size and gender imbalance. Future research should include larger and more diverse samples, conduct comparative studies across countries, and further develop conceptual and measurement frameworks for QQ in the South Korean context.
Practical implications – The findings highlight the importance of addressing excessive workloads, supporting work–life balance, and reconsidering traditional workplace structures that may contribute to employee dissatisfaction and burnout among younger workers.
Originality/value – Existing research on QQ has predominantly focused on Western contexts. This study contributes to the literature by providing qualitative insights into how South Korean Gen Z employees perceive and experience QQ within a distinct cultural and organisational setting.
