Contributions published at Department of Informatics (Burkhard Stiller)

The battlefield has evolved into a mobile and dynamic scenario where soldiers and heterogeneous military equipment exchange information in real-time and wirelessly. This fact brings to reality the Internet of Battlefield Things (IoBT). Wireless communications are key enablers for the IoBT, and their management is critical due to the spectrum scarcity and the increasing number of IoBT devices. In this sense, IoBT spectrum sensors are deployed on the battlefield to monitor the frequency spectrum, transmit over unoccupied bands, intercept enemy transmissions, or decode valuable information. However, IoBT spectrum sensors are vulnerable to heterogeneous cyber-attacks, and their accurate detection is an open challenge in the literature. Thus, this paper presents SpecForce, a security framework for IoBT spectrum sensors based on device behavioral fingerprinting and ML/DL techniques. SpecForce considers heterogeneous data sources to detect the most dangerous and recent cyber-attacks affecting IoBT spectrum sensors, such as impersonation, malware, and spectrum sensing data falsification attacks. To evaluate the SpecForce detection performance, it has been deployed on 25 real spectrum sensors, and results show almost perfect detection for the three cyber-attack families previously mentioned.

This thesis investigates the data privacy preservation capabilities of Federated Learning (FL), specifically focusing on Centralized Federated Learning (CFL) and Decentralized Federated Learning (DFL) settings. Despite their existing data privacy advantages, both CFL and DFL have been shown to be vulnerable to adversarial attacks, including Membership Inference Attacks (MIA). This thesis compares the data privacy-preserving capabilities of CFL and DFL, trained on MNIST, FashionMNIST, and CIFAR-10, against White-Box and Black-Box MIA across various performance metrics. Furthermore, the most commonly used defense techniques used against MIA are discussed, such as Differential Privacy (DP), Regularization, and Knowledge Distillation. The findings suggest that FL models generally provide better data privacy than ML models, with CFL being the best data privacy preserving federation against shadow models using binary classifier-based MIA and DFL models with a fully connected network topology, showing strong resistance against MIA using a prediction-based classifier. This work offers valuable insights into the data privacy-preserving abilities of CFL and DFL in different scenarios and underlines the importance of further research in the domain of data privacy in collaborative ML.

Adherence to therapy is a significant issue when treating patients who suffer from adiposity. To improve therapy success, modern technology can be used to assist in increasing patients' adherence to therapy. One such technology is the Digital Companion application, which includes a mobile app for patients to track their therapy progress and a web interface for doctors to access further analytics on their patients' ongoing therapy outcomes. This bachelor thesis investigates how we can contribute to the Digital Companion application by developing a Natural Language Processing engine that analyzes the journal entries written by patients in the Digital Companion mobile application. The analysis we provide consists of a topic-based sentiment analysis-oriented algorithm. Through this approach, our aim is to identify which aspects of the therapy are going well for the patient and which are not. This may not always be apparent to the doctor due to their limited time and resources when preparing for a specific patient's consultation and their lack of oversight of what occurs between consultations. With our proposed model, we achieved better precision and recall scores than other industry-leading models when evaluating the patients' data, demonstrating the effectiveness of our approach for the task at hand.

The importance of fairness in algorithmic decisions and artificial intelligence has grown continuously over the last years and the research on it has expanded to include the study of fairness perception. At the same time, the influence of cultural background on perception of fairness in algorithmic decisions remains largely unexplored in the do¬mains of cross-cultural fairness studies and fairness in artificial intelligence. The purpose of this thesis is to conduct a 3 (country, independent) x 2 (scenario, independent) x 3 (strategy, dependent) factorial design preliminary experiment to find out whether the perception of fairness in contexts of algorithmic decisions differs across cultures. To assess this, a survey was developed, detailing the use of an algorithm for decision making in a school admission and in a loan approval scenario, and how the employment of different fairness notions (independence, separation, and sufficiency) would influence those decisions. The survey was distributed online to 300 participants, 100 each per country (Germany, South Africa, and the United States of America), who were asked to rank the three fairness notions from fairest to least fair per scenario. The resulting data was analysed based on mean rank per strategy and rank frequency. Statistical tests employed to prove significance were not applicable to the data type at hand – neither ordinal, nor ranked. For further analysis, the 3 x 2 x 3 matrix was extended by the independent variables gender, age, and openness towards AI. The more independent variables were introduced to make specific statements, the more diverse the observed tendencies were. Nevertheless, results show a clear overall preference for one strategy as the fairest: separation. Except for scenario 1 in South Africa, where independence was ranked as fairest. The required insights on fairness perception in South African culture to explain this divergent tendency are not present in current literature on cross-cultural fairness.

Chronic diseases are a worldwide problem. However, many cases could be prevented or treated by lifestyle changes. The use of mobile health technologies like health apps offers new opportunities to support treatments. This thesis examines how the journal of such a health app affects the patient-doctor collaboration. The patients and doctors make treatment agreements about the behavior of patients in the phase between consultations. The patients have the agreements stored on the health app and can make notes about their execution. The doctors can review the notes in the journal before the consultation. They then discuss the journal with the patient in the consultation. In this thesis the interviews conducted after the follow-up consultation are analyzed. The patients report predominantly positive effects on the patient-doctor collaboration. For example, patients report that the journal had positive effects on the patient-doctor relationship and that they could collaboratively create a treatment plan that they believe to be sustainable. They also report that the doctor’s influence affected the motivation of the patients positively through the journal in the phase in between consultations. The insights about the importance of the patient-doctor relationship were used to extend an existing theoretical model.

This work focuses on the analysis and evaluation of methods for efficient indexing and querying of motion videos in multimedia databases. In particular, the software stack vitrivr is considered for the representation of a multimedia database. The emphasis is on investigating the feasibility of implementing an index for motion data from the motion videos in one of the components of the software stack. The index should then enable a more efficient query. In this context, the use of technologies such as OpenPose for extracting motion data from video sequences and Dynamic Time Warping (DTW) for analyzing and categorizing the extracted data is discussed based on previous work. The work further illuminates the application of local segmentation as a means of improving the accuracy and efficiency of a similarity assessment based on DTW. Finally, two possible implementation approaches for the index are proposed and analyzed: one that builds on the existing architecture of the retrieval engine Cineast, and another that suggests a direct integration of DTW functionality into Cottontail DB.

Many variants of (delegated) Proof-of-Stake blockchains have been introduced since 2016, among which Cardano, Algorand, and Tezos have been more widely adopted. Each of these blockchains has attempted to address the scalability and consumption issues of Proof-of-Work (PoW) blockchains from different angles. However, the economic incentives of these blockchains have raised the concern about the decentralisation level of key players, such as validators, and stakeholders, in those “expected to be decentralised” networks. In this work, we conduct an analysis of Cardano's activity and wealth concentration. For the first time, we perform a clustering of the addresses that belong to the same wallet based on existing heuristics for UTXO-like blockchains. Moreover, an analysis of reward distribution on Cardano is presented. The analysis results of this work disclose objective insights into the Cardano blockchain, including the rewards distribution, stake balance distribution, and wealth concentration.

Inertial odometry is an attractive solution to the problem of state estimation for agile quadrotor flight. It is inexpensive, lightweight, and it is not affected by perceptual degradation. However, only relying on the integration of the inertial measurements for state estimation is infeasible. The errors and time-varying biases present in such measurements cause the accumulation of large drift in the pose estimates. Recently, inertial odometry has made significant progress in estimating the motion of pedestrians. State-of-the-art algorithms rely on learning a motion prior that is typical of humans but cannot be transferred to drones. In this work, we propose a learning-based odometry algorithm that uses an inertial measurement unit (IMU) as the only sensor modality for autonomous drone racing tasks. The core idea of our system is to couple a model-based filter, driven by the inertial measurements, with a learning-based module that has access to the thrust measurements. We show that our inertial odometry algorithm is superior to the state-of-the-art filter-based and optimization-based visual-inertial odometry as well as the state-of-the-art learned-inertial odometry in estimating the pose of an autonomous racing drone. Additionally, we show that our system is comparable to a visual-inertial odometry solution that uses a camera and exploits the known gate location and appearance. We believe that the application in autonomous drone racing paves the way for novel research in inertial odometry for agile quadrotor flight.

While there are many data collection and analysis tools for Ethereum - the largest smart contract blockchain by market capitalization, development of similar tools for other smart contract blockchains is lacking. Reasons for this are non-existent standards, changing specifications d ue t o rapid development, common usage of the off-chain storage, and lack of developers. One of such blockchains is Polkadot - a layer- zero blockchain featuring a single relay chain whose role is to secure smart contract transactions on multiple other parachains. In this paper we describe a data collection pipeline for Polkadot blockchain that we then use to perform an analysis of the bow-tie structure of its transfer network over time, with special emphasis on the role of nominators and validators in this structure. We find evidence that t he Polkadot ecosystem iss lowly maturing from a system dominated by nominators and validators, both of which require some technical skill as well as willingness to bond sufficient amount of funds, into a system increasingly populated by regular users using the financial services of Polkadot.

The rapid development of the “sharing economy” enables the effective and efficient coordination, acquisition, distribution, and sharing of many kinds of different resources. Beyond the well-known sharing economy services such as Airbnb and Uber, an increasing number of local sharing initiatives have established online platforms and services to facilitate access to the shared resources within their local communities. With the automation and complexity of digital tools and platforms, and the specific challenges of online sharing communities, users’ trust and reliance become increasingly critical for successful use and adoption. In our qualitative study in collaboration with two industry partners (a local sharing community and a large infrastructure provider in Switzerland), we unpack various perspectives of interpersonal trust in the community and the systems trust of the supporting technologies. On this basis, we elicited a set of design opportunities for future platforms in the context of sharing economy.

Recent regulatory changes have enabled NCAA student-athletes to proft from their name, image, and likeness (NIL), departing from previous policies requiring those athletes to maintain their amateur status. However, despite the changes, it is unlikely that all the approximately 500,000 NCAA student-athletes will proft from NIL contracts. Within this context, we study how to design a fair and inclusive solution that may help all student-athletes se- cure NIL fnancial resources. Following a design science approach, we defne design requirements after interviewing student-athletes. Subsequently, we derive three design principles: inclusiveness, fairness, and transparency. Thereafter, we suggest a blockchain-based artifact that satisfes all design principles. Our idea lies in designing collectibles as non-fungible tokens (NFTs) that pay diferent roy- alties whenever a transaction (purchase or exchange) happens in diferent markets (primary or secondary). Finally, we evaluate our solution by discussing its features with current student-athletes.

Semantic segmentation is a growing topic in high-resolution remote sensing image processing. The information in remote sensing images is complex, and the effectiveness of most remote sensing image semantic segmentation methods depends on the number of labels; however, labeling images requires significant time and labor costs. To solve these problems, we propose a semi-supervised semantic segmentation method based on dual cross-entropy consistency and a teacher–student structure. First, we add a channel attention mechanism to the encoding network of the teacher model to reduce the predictive entropy of the pseudo label. Secondly, the two student networks share a common coding network to ensure consistent input information entropy, and a sharpening function is used to reduce the information entropy of unsupervised predictions for both student networks. Finally, we complete the alternate training of the models via two entropy-consistent tasks: (1) semi-supervising student prediction results via pseudo-labels generated from the teacher model, (2) cross-supervision between student models. Experimental results on publicly available datasets indicate that the suggested model can fully understand the hidden information in unlabeled images and reduce the information entropy in prediction, as well as reduce the number of required labeled images with guaranteed accuracy. This allows the new method to outperform the related semi-supervised semantic segmentation algorithm at half the proportion of labeled images.

Digital technologies amplify and change societal processes. So far, society and intellectuals have painted two extremes of viewing the effects of the digital transformation on democratic life. While the early 2000s to mid-2010s declared the "liberating" aspects of digital technology, the post-Brexit events and the 2016 US elections have emphasized the "dark side" of the digital revolution. Now, explicit effort is needed to go beyond tech saviorism or doom scenarios. To this end, we organized the Dagstuhl Seminar 22361 "Challenges and Opportunities of Democracy in the Digital Society" to discuss the future of digital democracy. This report presents a summary of the seminar, which took place in Dagstuhl in September 2022. The seminar attracted scientific scholars from various disciplines, including political science, computer science, jurisprudence, and communication science, as well as civic technology practitioners.

null

The measurement of credibility for Twitter content has gained significant attention due to the difficulty in verifying the accuracy of posts, particularly those made by users who identify themselves as experts by including titles such as Dr.’ or M.D.’ in their display name. This study aimed to investigate three research questions. First, we assessed the credibility of users who display qualified titles on Twitter. Next, we analysed the types of viewers who are most susceptible to the influence of such users, and finally we proposed strategies that can be used by actual ‘Dr.’ titled users to enhance their credibility on the platform. To gather data, a between-subject experiment and a survey were designed and conducted. The results indicate that users with professional titles in their display names are perceived as more credible than those without such titles. Additionally, the study found that individuals who have never used Twitter before are the most impacted by Twitter content. Our study suggests that real ‘Dr.’ titled users can increase their credibility by including a relevant bio in their profile and by including paper links in their tweets. By doing so, these users can more effectively persuade the public of their expertise.

In our work, we propose a novel automated radiology-specific evaluation metric that can be used for evaluating the performance of machine-generated radiology reports. We utilize the existing successful COMET metric architecture, which we adapt and optimize for use in the radiology domain. Using this architecture, we train and publish four medically-oriented model checkpoints using various combinations of encoders and corpora of radiology reports. One of the model checkpoints is trained using RadGraph, a radiology knowledge graph, and the thereof-derived RadGraph F1 and RadCliQ scores are integrated into our contributed parallel corpora to enhance their quality. Our results show that the developed metric exhibits a moderate to high correlation with established metrics such as BERTscore, BLEU, and S_emb score, indicating its potential effectiveness as a radiology-specific evaluation metric.

Dependency parsers tend to struggle with parsing transcribed spoken language as they are trained on properly structured, written text. Spoken language lacks the structure of properly written text and exhibits typical phenomena like disfluency, repetition, and truncation of words and sentences. This research examines the problem of parsing spoken language for Swiss German audio transcripts from ArchiMob corpus. Swiss German, an umbrella term for the German (Alemannic) dialects spoken in Switzerland, lacks orthographic and grammatical standardization, shows a high degree of variation among the various dialects and differs substantially from Standard German. The lack of standardization is due to the situation of diglossia in Switzerland. As Swiss German is mainly an oral language or restricted to informal writing, many resources lack structure and exhibit a high variability in terms of morphology, spelling and vocabulary. The combination of variation in Swiss German, its lack of standardization and the unstructuredness of spoken language render parsing transcribed Swiss German challenging. Accordingly, pre-trained (German) dependency parsers struggle with Swiss German audio transcripts and little data is available to train them. This research tackles the problem of parsing spoken language by re-segmenting Swiss German audio transcripts into sentence-like units (SLUs) and examines the impact of re-segmentation on dependency parser performance. Therefore, our experiment setup includes two evaluation steps, one for re-segmentation and one for dependency parsing. We frame the re-segmentation as a binary classification task aiming to predict tokens marking an SLU-boundary. For this purpose, we fine-tune a pre-trained German Bert model to predict such boundaries. These predicted SLU-boundaries are used to re-shape the input for the dependency parser. We show that the re-segmentation into SLUs leads to an improvement of the Labeled Attachment Score (LAS) over a baseline. Moreover, we demonstrate that the performance in the SLU-boundary classification task correlates with the parser performance. To engage in such a supervised learning setting, a test set composed out of roughly 200 SLUs was manually created and annotated with dependency labels for the two folded evaluation. With our work, we contribute to processing spoken Swiss German by showing a way of inducing more structure.

Gender bias appears in many neural machine translation models and commercial translation software. The problem is well known and efforts to reduce such discriminatory tendencies are underway. But gender bias is still not fully solved. This work utilizes a controlled text generation method, Future Discriminators for Generation (FUDGE), to reduce the so-called Speaking As gender bias emerging when translating from English to a language that openly marks the gender of the speaker. The model is evaluated with BLEU and MuST-SHE, a novel gender translation evaluation method. The results demonstrate improvements in the translation accuracy of the feminine terms.

Financial markets are complex and constantly evolving systems, where investors make decisions based on market conditions, company performance, and global economic trends. However, recent studies suggest that peer effects can also play a significant role in shaping investment decisions. Peer effects refer to the influence that one's peers have on their decision-making, and in the context of financial decision-making, can cause investors to follow trends in herding behavior. This influence process can result in cascading behavior, where the actions of a few investors can trigger a chain reaction of buying or selling, leading to significant price movements. The impact of peer effects has been amplified by social networks that have revolutionized the way we communicate and share information. In this thesis, we investigate the role of peer effects in financial markets and their impact on cascading behavior. Using a real-life evolving socio-financial network, we aim to quantify the extent to which individual investors influence the generation of cascading behavior, with a particular focus on the spatio-temporal features of individual investors within the network. We formulate a prediction task that forecasts the influence of individual users by utilizing various centrality measures and time-aware node embeddings. We evaluate the effectiveness of these centrality measures and time-aware node embeddings in predicting the influence of users in generating cascades of trades through the network. Our study contributes to a better understanding of the spatio-temporal factors that facilitate cascading behavior in financial markets, highlighting the need to understand their impact in various contexts, including real-life socio-financial networks.

Modern Code Review is an essential step of software development processes in industrial settings and open-source projects. It is usually supported by various tools to help reviewers during the process. Nonetheless, a significant part of the review time is still spent on understanding submitted changes. The challenge of understanding code changes could be improved by new tools designed for change authors to help them create more reviewable changes. In this study, I collected information on different aspects relevant to the design of such tools, including their responsibilities and the associated implementations. I present Cres, a tool designed for identifying oversized commits and helping developers divide them into smaller commits. Cres was implemented following two different approaches, resulting in a web application and a pair of Git hooks. Both approaches were evaluated in interviews with expert developers to provide ideas and advice for the design of future tools.