Contributions published at Department of Informatics (Burkhard Stiller)

Government as a Platform (GaaP) is a promising approach to the digital transformation of the public sector. In practice, GaaP is realized by platform-oriented infrastructures. However, despite successful examples, the transformation toward platform-oriented infrastructures remains challenging. A potential remedy is the analysis of existing public infrastructure regarding its platform orientation. Such an analysis can identify the gaps to an ideal platform-oriented infrastructure and, thus, support the transformation toward it. We follow the design science research methodology to develop a four-dimensional analysis method. We do so in three iterations, and, after each iteration, evaluate the method by its application to infrastructures in practice. With regard to theory, our results suggest extending GaaP conceptualizations with a specific emphasis on platform principles. With regard to practice, we contribute an analysis method that creates proposals for the improvement of infrastructures and, thus, supports the transformation toward GaaP.

News recommender systems are an increasingly popular field of study that attracts a growing interdisciplinary research community. As these systems play an essential role in our daily lives, the mechanisms behind their curation processes are under scrutiny. In the area of personalized news, many platforms make design choices driven by economic incentives. In contrast to such systems that optimize for financial gain, there can be norm-driven diversity systems that prioritize normative and democratic goals. However, their impact on users in terms of inducing behavioral change or influencing knowledge is still understudied. In this paper, we contribute to the field of news recommender system design by conducting a user study that examines the impact of these normative approaches. We a.) operationalize the notion of a deliberative public sphere for news recommendations, show b.) the impact on news usage, and c.) the influence on political knowledge, attitudes and voting behavior. We find that exposure to small parties is associated with an increase in knowledge about their candidates and that intensive news consumption about a party can change the direction of attitudes of readers towards the issues of the party.

The Creator Economy faces concerning levels of unfairness. Content creators (CCs) publicly accuse platforms of purposefully reducing the visibility of their content based on protected attributes, while platforms place the blame on viewer biases. Meanwhile, prior work warns about the “rich-get-richer” effect perpetuated by existing popularity biases in recommender systems: Any initial advantage in visibility will likely be exacerbated over time. What remains unclear is how the biases based on protected attributes from platforms and viewers interact and contribute to the observed inequality in the context of popularity-biased recommender systems. The difficulty of the question lies in the complexity and opacity of the system. To overcome this challenge, we design a simple agent-based model (ABM) that unifies the platform systems which allocate the visibility of CCs (e.g., recommender systems, moderation) into a single popularity-based function, which we call the visibility allocation system (VAS). Through simulations, we find that although viewer homophilic biases do alone create inequalities, small levels of additional biases in VAS are more harmful. From the perspective of interventions, our results suggest that (a) attempts to reduce attribute-biases in moderation and recommendations should precede those reducing viewers’ homophilic tendencies, (b) decreasing the popularity-biases in VAS decreases but not eliminates inequalities, (c) boosting the visibility of protected CCs to overcome viewers’ homophily with respect to one fairness metric is unlikely to produce fair outcomes with respect to all metrics, and (d) the process is also unfair for viewers and this unfairness could be overcome through the same interventions. More generally, this work demonstrates the potential of using ABMs to better understand the causes and effects of biases and interventions within complex sociotechnical systems.

A central question in robotics is how to design a control system for an agile mobile robot. This paper studies this question systematically, focusing on a challenging setting: autonomous drone racing. We show that a neural network controller trained with reinforcement learning (RL) outperformed optimal control (OC) methods in this setting. We then investigated which fundamental factors have contributed to the success of RL or have limited OC. Our study indicates that the fundamental advantage of RL over OC is not that it optimizes its objective better but that it optimizes a better objective. OC decomposes the problem into planning and control with an explicit intermediate representation, such as a trajectory, that serves as an interface. This decomposition limits the range of behaviors that can be expressed by the controller, leading to inferior control performance when facing unmodeled effects. In contrast, RL can directly optimize a task-level objective and can leverage domain randomization to cope with model uncertainty, allowing the discovery of more robust control responses. Our findings allowed us to push an agile drone to its maximum performance, achieving a peak acceleration greater than 12 times the gravitational acceleration and a peak velocity of 108 kilometers per hour. Our policy achieved superhuman control within minutes of training on a standard workstation. This work presents a milestone in agile robotics and sheds light on the role of RL and OC in robot control.

In the non-profit sharing economy context, an increasing number of resource sharing collectives and organizations (e.g., libraries of things) and peer-to-peer grassroots sharing initiatives leverage underutilized household resources (e.g., tools) to optimize their shared use for the benefit of their local communities. However, a number of social-technical challenges prevent the endurance and growth of such initiatives. Prior research highlighted the specific difficulties related to poor visibility of members’ activities and often high social barriers that hinder interactions among neighbors and strangers. In our prior work, stemming from our continuous engagement with one local sharing community in Switzerland over several years, through fieldwork, interviews, and co-creation studies, we elicited a set of design opportunities to address the emergent community’s challenges. Based on these design considerations, we developed Züri teilt, a mobile application to facilitate resource sharing practices among neighbors aligning with the slow, temporal, and gradual nature of their relationships.

Large language models (LLMs) like ChatGPT recently gained interest across all walks of life with their human-like quality in textual responses. Despite their success in research, healthcare, or education, LLMs frequently include incorrect information, called hallucinations, in their responses. These hallucinations could influence users to trust fake news or change their general beliefs. Therefore, we investigate mitigation strategies desired by users to enable identification of LLM hallucinations. To achieve this goal, we conduct a participatory design study where everyday users design interface features which are then assessed for their feasibility by machine learning (ML) experts. We find that many of the desired features are well-perceived by ML experts but are also considered as difficult to implement. Finally, we provide a list of desired features that should serve as a basis for mitigating the effect of LLM hallucinations on users.

Rankings serve to structure large datasets. They are an integral part of decision-making in various fields. To address the complex task of interpreting and comparing ranking algorithms and the results they produce, we created a Visual Analytics (VA) tool called RankViz. This prototype allows users to visualize and explore the output of various ranking algorithms and includes multiple metrics to assess the quality and differences of the rankings. The pairwise comparison data we used to construct the rankings was collected in a previous study by Barth et al. This thesis reviews different ranking algorithms, details the functionality of RankViz, demonstrates its utility with usage scenarios, and discusses potential future work in this field.

Recommender systems have become integral in guiding users through the overwhelming abundance of online content. As these systems assume an ever-increasing role in shaping user decisions and preferences, there is a growing demand for clarity in their decision-making processes to instill trust. Recommendation algorithms with a high degree of accuracy such as matrix factorization are highly regarded and widely adopted. Nonetheless, these algorithms tend to exhibit high complexity in their logic and architecture, rendering them challenging to explain to end-users. This issue has been recognized and many tools have presented possible solutions. Many of the implemented approaches, however, have demonstrated shortcomings due to disregarding some user-centered properties or overly concentrating on unraveling the underlying algorithmic intricacy. This work presents MFExplain, an innovative tool for explaining movie recommendations generated with matrix factorization. The tool aims to explain recommendations by relying on the provision of intuitive justifications. Leveraging interactivity and cutting-edge dimensionality reduction techniques enables the tool to also encourage exploration, allow user feedback, and foster many desirable recommender system properties that enrich the user experience.

With the surge in the volume and dimensionality of large image datasets across fields such as medicine, manufacturing, and quality monitoring, there is an increased emphasis on efficiently curating these datasets. This design study explores the challenges associated with labeling and exploring large, high-dimensional, and unlabeled image datasets. Traditional tools prioritize either data visualization using techniques like dimensionality reduction or labeling automation using AI learning mechanisms. This binary focus often comes at the cost of extensive labeling functionalities or comprehensive overviews, since user interaction is reduced. This research bridges this gap by introducing DaedalusData, an interactive visual analytics approach that combines meaningful visual exploration with efficient labeling and intuitive feedback loops. DaedalusData presents an interactive platform that enables pattern and anomaly exploration, efficient image labeling by integrating metadata, and preliminary steps toward labeling automation. The tool was developed alongside domain experts and built for a dataset containing particle contamination in consumables at Roche Diagnostics. As a design study this thesis, solved a real-world problem, through close collaboration with domain experts. The study posits that merging interactivity, human expertise, and automated processes offer a promising direction for managing large image datasets, with DaedalusData serving as a foundational step.

Tuberculosis (TB) is an infectious disease caused by the bacterium Mycobacterium tuberculosis, which is one of the leading causes of death worldwide. Various Deep Convolutional Neural Network models have gained popularity to help during the TB screening process by detecting patients with active Tuberculosis from their Chest X-Rays. To help with further advancing the research, a new publicly available dataset, TBX11K, has been used to increase the number of samples during training for existing replaceable state-of-the-art models. In the first step, the model’s performance was evaluated to see if an improvement through the addition of more TB-related data was observable. It was shown that state-of-the-art replicable binary classifier models could further be improved through the inclusion of more data. Further, there is a lack of focus on generating and evaluating explanations for such models. The preferred methods currently are saliency mapping techniques such as Grad-CAM, to generate visual explanations based on the model’s decision-making process, by overlaying heatmaps over the Chest X-Rays. The selected TBX11K dataset includes ground truth bounding box labels, which makes it possible to evaluate if the visualisations were correct. There are various evaluation metrics to evaluate the faithfulness and localisation performance of the saliency mapping techniques according to ground truth labels. Two of them have been identified to be useful, namely RemOve and Debias, and Proportional Energy. RemOve and Debias was used to observe if there is one universal saliency mapping technique that performs well for all models for the task of active Tuberculosis detection. Further, based on these two metrics, a new metric was proposed, ROAD-Normalised PropEng Average, to measure the overall best-performing model and Saliency Mapping Technique combination. From the evaluation with RemOve and Debias, it was concluded that there does not seem to be a universal saliency mapping technique that performs well on all model architectures for the de- tection of active Tuberculosis. Thus, it is recommended to always consider the underlying model before choosing the optimal saliency mapping technique. Further, through the use of the ROAD-Normalised PropEng Average, it was concluded that one model in combination with a saliency mapping technique offered the best trade-off between faithfulness and correctness of the visualisations. This was the multi-label DenseNet-121 model with Eigen-CAM. To obtain accurate clas- sifications of active Tuberculosis with explainable and correct visualisations, it is recommended to use this model and visualisation technique combination.

A problem with modern deep learning recognition systems is that they often respond to stimuli of an unknown class overly confident, but wrong. Open-set recognition highlights this behavior and provides evaluation methods to estimate the generalization capability of models beyond the classic train/test set split. In this thesis, we incorporate a Radial Basis Function (RBF) layer into deep convolutional networks to model the deep feature distribution. We evaluate such networks on standard open-set evaluation protocols and compare their performance with standard Softmax classification models. Additionally, we utilize negative training samples and compare with the Entropic Open-Set Loss Softmax extension. We show that standard deep RBF network with Gaussian activation functions does not outperform Softmax based methods in open-set recognition. We extend the RBF network in two ways, which both show increased open-set recognition performance over the baseline RBF network. Based on these results we conjecture that solely using an RBF layer for the classification sub-system of a deep neural network might not be sufficient to solve the open-set recognition problem.

Many users tend to seek information about specific countries on TikTok before planning international travel. To investigate the objectivity and authenticity of national information on TikTok, we examine the national images presented on TikTok for 12 popular tourist countries from different continents. We utilize web scraping techniques to collect TikTok data for these countries, including descriptive data of videos, watermark-free video contents, comments, etc. Subsequently, we conduct separate analyses of descriptions, video contents, and comments of TikTok videos related to each country. Additionally, we design a survey questionnaire to gather user perceptions of various national images on TikTok. Ultimately, through the analysis of TikTok data and survey responses, we identify consistent trends and specific themes in the descriptions, video contents, and comments of TikTok videos related to specific countries. For instance, TikTok videos related to Argentina predominantly revolve around the theme of football, while those related to Italy and Spain primarily focus on travel and food. Furthermore, users' impressions of specific countries on TikTok closely align with the national images presented on the platform.

This research utilized eye tracking technology to gain a better comprehension of the code review process. We collected data from 14 participants, ranging from inexperienced Java users to experienced Java developers and code reviewers with more than a decade of experience. By analyzing the eye tracking data, we were able to identify differences in attention patterns based on the size of the code changes and the focus on various code elements. Notably, smaller code changes received more detailed attention to specific code elements than larger ones. Our results provide useful information that can be used to improve code review processes and developer training.

The rise of Blockchain technology and cryptocurrencies has enabled the creation of decentralized alternative payment systems, without the need for third party financial institutions. However, this decentralization and pseudo-anonymity have also facilitated the emergence of darknet markets (DNMs) offering illicit products. This study introduces the"grayscale diffusion framework", with the aim of modeling and understanding the propagation of dark assets in the Bitcoin network. Formulated and implemented with a combination of on-chain and off-chain data, this approach utilizes address clustering, haircut tainting, and community partitioning to offer a unique analysis perspective on dark asset proliferation across the Bitcoin blockchain. The framework uncovers interesting patterns in the assortative nature of Bitcoin transactions based on the darkness level of assets, pointing to the existence of non-random clusters and communities that facilitate dark asset diffusion. Our research not only addresses key questions related to the effective modeling and tracking of tainted asset flow, but also provides valuable insights. Keywords: Bitcoin, Address Clustering, Darknet Marketplaces, Gray-scale Diffusion, User Transaction Networks.

This thesis addresses the implementation of concepts outlined in FIFA's Enhanced Football Intelligence (EFI) document through an open-source library, filling the gap with accessible implementations for these concepts. The EFI document provides descriptions for various metrics related to football performance analysis used in the FIFA World Cup 2022. Existing packages in football analytics do not fully incorporate the latest methodologies used in the FIFA World Cup 2022, essential for the creation of a source that aligns with FIFA's definitions. The implemented concepts cover possession control, phases of play, ball recovery time, line breaks, receptions behind midfield and defensive lines, defensive line height and team length, team shape, final third entries, pressure on the ball, forced turnovers, and expected goals (xG). Utilizing the explanations of these concepts, the thesis formulates a main approach and involves refinements. The level of stability varies, with methods that incorporate fewer heuristics tending to be more stable, while those that rely on a greater number of heuristics tend to be less stable. However, during implementation, limitations were encountered, including the lack of technical details and absence of FIFA's resources regarding the technology they have employed. Specifically, the lack of heuristics mentioned in the definitions of the concepts was a notable gap. Challenges were also observed, such as specific matches that are labeled as outliers due to their performance in distinct concepts. Despite these limitations and challenges, the implementation overall offers stable and accurate performance, aligned with FIFA's outcomes. In future work, these limitations can be addressed through a comprehensive approach. Firstly, revisiting the concepts with additional information regarding their descriptions will enhance the understanding of the underlying factors. Secondly, the expansion of datasets will not only provide a broader foundation for analysis but also improve the heuristics employed, leading to enhanced accuracy and stability of the outcomes. Additionally, the application of advanced technologies, similar to those employed by FIFA, can significantly contribute to improving the reliability and effectiveness of the results. By considering these avenues, future research can overcome the identified limitations. This thesis contributes to advancing football performance analysis by addressing these challenges and provides a valuable resource for researchers, analysts, and football enthusiasts seeking to reproduce FIFA's match reports and gain insights into football performance. Keywords: Enhanced Football Intelligence, FIFA, open-source implementation, football performance analysis, sports analytics.

Representing 3 dimensional scenes as Neural Radiance Fields (NeRF) has shown impressive results for novel view synthesis. Generalizable and dynamic variations of NeRF have been studied extensively producing photorealistic results. However, a generalizable and dynamic NeRF remains a very challenging problem. An effective solution to this problem requires a large and diverse dataset portraying complex subject motion. In this work we provide an end-to-end framework for generating high-quality synthetic datasets with complex and realistic human motion tracked by multiple cameras moving along pseudo random trajectories as well as multiple static cameras.

In recent years, small and midsize enterprises (SMEs) have become increasingly reliant on technology, but lag in terms of investment into cybersecurity. This renders them vulnerable to malware attacks, which are increasingly targeting companies rather than individuals, with great economic impact. This project proposes and implements a prototype tool, which allows for machine learning models to be trained, stored, and tested within the SecBox sandbox environment. Both classification and anomaly detection models are implemented through Scikit-learn, in order to provide predictions about known malware types (binary and multiclass classification), as well as detecting the presence of unseen malware in real-time during the SecBox execution. The models are trained using the system call and resource usage file execution logs available from the SecBox, which are transformed into suitable formats using frequency-based and sequence-based data preprocessing. Model reproducibility is ensured by generating configuration files with references to the random seeds, the datasets used in training, as well as other model parameters, which can be used to re-train the same model. To evaluate and compare model performance, each model type is tested in a realistic scenario of the execution of Monti ransomware within the SecBox, creating a confusion matrix as well as calculating the accuracy, precision, recall and F1-score metrics based on the model predictions. The system call classifier models are shown to have the best performance when classifying Monti malware samples, and the project is concluded by specifying several relevant research areas to be investigated further.

This bachelor-thesis tries to incorporate Software Defined Networking (SDN) mechanisms into a Long Range (LoRa) Internet of Things (IoT) mesh. Typically, devices used in LoRa-based wireless sensor networks (WSN) are limited in range. Therefore, in a scenario where a significant amount of nodes would be out of range of the LoRa Wide Area Network (LoRaWAN), deploying a mesh topology is a simple yet effective way to connect far away nodes using multi-hop communication. SDN, on the other hand, aims to improve network performance by analyzing the network and applying smart optimization algorithms. The hardware used in this thesis, are nodes being a Raspberry Pi, a popular choice inIoT, and the E32-868T20D LoRa Shield, a budget-friendly option to adapt LoRa technology. The software is implemented in Java, a programming language that promotes human-readability in code and benefits from decades of experience in practical software development. While LoRa networks with mesh topologies have already been subject of previous research using various devices and programming languages, the goal of this thesis is to test the effectiveness of SDN-based mechanisms in improving a LoRa mesh network and finally providing a user-friendly API as a service to other applications as a transmitter of data. Disclaimer: Neither the product, nor the analysis of SDN-based mechanisms have reached a state of success.

The Internet-of-Things (IoT) is becoming increasingly integral to modern living, offering a wide range of applications across diverse devices employing a multitude of different operating systems (OS). While Linux is the most prevalent OS in the IoT landscape, its resource requirements often prevent its use on less powerful, cost-efficient devices like those in the ESP32 family of microcontrollers (MCUs). The goal of this thesis is therefore to explore the feasibility of porting Linux to ESP32 devices, motivated by both economic and IoT standardization incentives. A cost analysis reveals an approximately 80% reduction in expenses when using the ESP32-WROVER-IE module compared to the Raspberry Pi Zero W. A tool chain was constructed to compile a Linux kernel image, which was successfully ported to an ESP32-S3-DevKitC-1 board. Initial evaluations indicate that the ported system offers basic functionality suitable for IoT tasks, although certain limitations currently restrict its practical utility.

Nowadays, companies and organizations invest in cybersecurity more and more as they are operating with digital information systems. Cyber risk management presents a well-defined path toward the management of critical assets, threats, and countermeasures. Within cyber risk management, threat modeling is a structured process to identify potential threats, and in this process, it is significant to evaluate each threat and estimate its potential impacts. Although threat modeling methodologies have been developed in depth, most of them focus on threat identification in di↵erent contexts, while how to quantify their impact for further inspection is less discussed. This thesis works on designing a framework to fill in this gap. The main outcome of this thesis is a framework that guides users to evaluate and quantify cyber threats in business contexts. The framework integrates applicable business impacts, calculates and visualizes the impacts of cyber threats, providing users with an intuitive picture of cyber threats analysis in the view of business. The prototype is well developed and properly evaluated, and the usability of the prototype is of satisfaction.