Themen

Hintergrund und Ziel ============= Analyse des aktuellen Stand der Forschung zu den FAIR Principles in Cyber Security. Aufgabenstellung ========== * Systematische Recherche des aktuellen Stands der Forschung * Analyse der aktuellen Themen * Gapanalyse * Konzepterstellung * Interpretation der Resultate

Ziel der Arbeit ist es ein Security Testing Framework für Kollaborative Roboter zu entwickeln und eigene Tests durchzuführen. Aktueller Stand der Forschung ist eher übersichtlich und zeigt erste Möglichkeiten auf. Es gibt jedoch bereits eine Fülle von Werkzeugen, die z.B. Penetration Testing unterstützen. Jedoch gibt es in dem Bereich wenig Erfahrung im Bereich der Kollaborativen Roboter. Im Rahmen der Arbeit soll der aktuelle Stand der Technik evaluiert, ein Rahmenwerk für Security Testing von Kollaborativen Robotern erarbeitet (Welche Angriffsvektoren gibt es? Welche Technologien? Etc.) Mindestens zwei Anwendungsfälle mit offline Simulatoren werden erarbeitet und durchgeführt. Diese sollen die Herausforderungen des aktuellen Stands hervorheben und mögliche Lösungsansätze beschreiben. Referenz: Hollerer, S., Fischer, C., Brenner, B., Papa, M., Schlund, S., Kastner, W., ... & Zseby, T. (2021). Cobot attack: a security assessment exemplified by a specific collaborative robot. Procedia Manufacturing, 54, 191-196.

As the demand for customized products and efficient use of resources has grown significantly, especially among Small and Medium-sized Enterprises (SMEs), these companies are approaching collaborative robots (cobots) for automating tasks and adapting to these demands. The cobots prioritize the safety of human operators and offer high flexibility. As the number of such robots grows, the need to manage their coordination and collaboration arises too. This presents new challenges for ensuring safety, security and coordination in shared spaces. Therefore, this study will contribute to addressing these challenges by providing a structured overview of the existing multi-human-cobot collaboration. In this context, a multi-human-cobot setup must include atleast two cobots or two humans along with the other, to realize our research. Description: • The thesis involves conducting a systematic literature review on the existing collaboration for multi-human-robot systems. • The goal is to provide an overview of how a multiple robotic setup coordinate with each other. • The focus is to summarize the existing research and practical approaches from other research, industries and deployed solutions, simplifying complex information in the survey. Learning Objectives: • Gain knowledge about cobots, their architectures and state-of-the-art solutions. • To learn how collaboration in robots is organized and coordinated. • Compare different collaboration approaches. • Identify common challenges and open questions in this direction.

Background ======================= As the collaborative robots (cobots) move beyond the horizon of single-human-robot interaction, into the domain of multiple actors (humans and robots) working in shared workspaces, thus, evaluating their performances through traditional metrics like cycle time, success rates, repeatability, etc. are insufficient to capture the complexity of a multi-human-robot collaboration (multi-HRC) system. The aim of this seminar is to explore and classify existing and new metrics that can be suitable for multiple-HRC systems. It will include a detailed review of existing literature from human factors, ergonomics, control theory, etc. along with the assessment of interaction quality, safety, fluency, coordination and other parameters in a shared setting. ======================= Learning Objectives: - Understand the role of metrics evaluation in multi-HRC performance and safety. - Conduct a literature review on metrics from robotics, human-machine interaction and cognitive science. - Categorize metrics in different dimensions. For e.g. fluency, safety, security, mutual awareness and trust. - Identify gaps in current evaluation methods for multi-HRC. - Propose novel candidate metrics for collaborative systems. Example references: 1. Saleh, J.A., Karray, F. (2011). Towards Unified Performance Metrics for Multi-robot Human Interaction Systems. In: Kamel, M., Karray, F., Gueaieb, W., Khamis, A. (eds) Autonomous and Intelligent Systems. AIS 2011. Lecture Notes in Computer Science(), vol 6752. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-21538-4_31 2. Lanssie Mingyue Ma, Martijn Ijtsma, Karen M. Feigh, and Amy R. Pritchett. 2022. Metrics for Human-Robot Team Design: A Teamwork Perspective on Evaluation of Human-Robot Teams. J. Hum.-Robot Interact. 11, 3, Article 30 (September 2022), 36 pages. https://doi.org/10.1145/3522581

Situation: The predictions created by machine learning models are increasingly used as decision support in organizations to reduce costs and increase profits. In high-stakes decision contexts, intrinsically interpretable machine learning (IIML) models are essential because they enable decision-makers to understand the model’s internal decision logic from data input to the prediction output and, ultimately, to trust its predictions. Problem: IIML algorithms learn patterns directly from the data they are trained on. Consequently, they may also learn spurious, incorrect, or misleading relationships present in the data. As a result, the intrinsic interpretations provided by these models, and communicated through graphical visualizations, may be distorted and inconsistent with established domain knowledge. Question: How to design a mechanism that aligns IIML models with domain knowledge? Solution: Development of a mechanism that aligns an IIML model with domain knowledge. For example, such a mechanism can be designed in two ways: i) as a post-hoc technique that adjusts model interpretations after model training or ii) as an integral component of the learning objective (e.g., the loss function) that adjusts model interpretations during model training. Literature: - Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. - Mao, L., Wang, H., Hu, L. S., Tran, N. L., Canoll, P. D., Swanson, K. R., & Li, J. (2024). Knowledge-informed machine learning for cancer diagnosis and prognosis: A review. IEEE Transactions on Automation Science and Engineering. - Wang, Z. J., Kale, A., Nori, H., Stella, P., Nunnally, M., Chau, D. H., Vorvoreanu, M., Vaughan, J. W., & Caruana, R. (2021). GAM changer: Editing generalized additive models with interactive visualization. arXiv preprint arXiv:2112.03245.

Situation: The predictions created by machine learning models are increasingly used as decision support in organizations to reduce costs and increase profits. In high-stakes decision contexts, intrinsically interpretable machine learning (IIML) models are essential because they enable decision-makers to understand the model’s internal decision logic from data input to the prediction output and, ultimately, to trust its predictions. Problem: Most IIML algorithms learn only correlations between features (e.g., a higher number of higher ice cream sales is associated with a higher number of drowning incidents) to maximize performance in prediction tasks. In doing so, underlying causal relationships between features (e.g., hot weather increases both the number of ice cream sales and the number of drowning incidents) are often ignored. Consequently, the interpretations provided by these models, and visualized through plots, may be less meaningful for decision-makers because they reflect correlations and proxy variables rather than the true causal drivers of the outcome. Question: How to design a method for causal and intrinsically interpretable machine learning? Solution: Development of a method that combines causal machine learning with IIML. For example, a possible method can consist of three phases: i) discovering causal structures, ii) integrating the identified causal structures into an IIML model, and training the adapted model to create predictions and interpretations that are both transparent and causally grounded. Literature: - Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. https://doi.org/10.1038/s42256-019-0048-x - Zapata Gonzalez, D. (2026). A step towards inherently interpretable causal machine learning models for decision support. In Proceedings of the 34th European Conference on Information Systems. AIS Electronic Library. https://aisel.aisnet.org/ecis2026/bus_analytics/bus_analytics/7 - Cui, P., & Athey, S. (2022). Stable learning establishes some common ground between causal inference and machine learning. Nature Machine Intelligence, 4(2), 110–115.