Contributions in Merlin

In this study, we explore the susceptibility of financial sentiment analysis to adversarial attacks that manipulate financial texts. With the rise of AI readership in the financial sector, companies are adapting their language and disclosures to fit AI processing better, leading to concerns about the potential for manipulation. In the finance literature, keyword-based methods, such as dictionaries, are still widely used for financial sentiment analysis due to their perceived transparency. However, our research demonstrates the vulnerability of keyword-based approaches by successfully generating adversarial attacks using the sophisticated transformer model, GPT-3. With a success rate of nearly 99% for negative sentences in the Financial Phrase Bank, a widely used database for financial sentiment analysis, we highlight the importance of incorporating robust methods, such as context-aware approaches such as BERT, in financial sentiment analysis.

Switzerland has become the latest country to put words into action working towards more sustainable finance. During the Point Zero Forum in Zurich, the three-day event connecting policy with technology, the new Swiss Green Fintech Network (GFN) was launched, aiming to boost the green digital finance ecosystem.

This paper explores the public value outcomes generated by collaborative platforms in public services through a comparative case study of 25 privately-owned digital platforms across various public service areas in Italy. The study confirms that collaborative digitally-enabled endeavors can have both positive and negative effects on public value, highlighting the challenges associated with multi-actor dynamics and the delicate balance between public interest and private gain in public services. The findings reveal implications related to user-centric platforms, including concerns about inclusivity, autonomy, decision-making abilities, and privacy infringement. Additionally, the study suggests that collaborative platforms alone do not enhance collaboration in public services unless accompanied by strong public governance that promotes interoperability through standardized frameworks, common templates, and data reuse.

null

With an aging population across the world, dementia and neurological diseases, such as Alzheimer's disease (AD), are on the rise. The disproportionate rise in AD cases in developing countries gives rise to a low-cost, robust way to diagnose early-onset AD. The EARDREAM project takes up the fight against AD in these countries using low-density electroencephalography (EEG) device, with the goal of developing a digital biomarker of early-onset AD. To make the collected health data accessible and enable large-scale analysis, there is a need for accessible, scaleable, secure solutions for EEG data analysis. This thesis presents the design and implementation of the Wondernap platform, a novel, cloud-based system dedicated to enhancing the current process of interacting with the data generated using the EARDREAM EEG device. The evolution of the platform through two iterative prototypes is described, highlighting the transition from an initial prototype to a scaleable, secure, cloud-based solution. The initial prototype laid the groundwork for a scaleable, modular, and transferable architecture capable of accounting for the unique requirements of the EARDREAM project, employing state-of-the-art technologies for EEG data analysis architectures. Deploying a Flask backend and Apache HBase for EEG data storage, with MongoDB for patient data, the first iteration validated its usability through a user evaluation, scoring 84/100 on the System Usability Scale. Building upon user feedback and stakeholder input, the second prototype accentuated the applicability of cloud computing to the current architecture, demonstrating its scalability and portability. Using infrastructure-as-code and incorporating Apache Phoenix, this prototype showcased enhancements in fault tolerance, security, and performance. In summary, the developed platform offers a fault-tolerant, scaleable, secure cloud architecture supporting a user-friendly frontend allowing its users to gain insights into the data generated using the EARDREAM portable EEG device.

Chatbots have experienced a significant surge in popularity in recent months, which can largely be attributed to the utilization of Large Language Models (LLMs). Among these platforms, ChatGPT has shown the fastest growth, amassing one million users within a only five days. This can be attributed to the inherent contextual understanding and impressive capabilities exhibited by LLMs, which continue to be explored. Recognizing the potential of chatbots, particularly their adaptability to custom interfaces, our aim is to develop a tailored assistant to help adults in corporate identity E-learning. Our pedagogical conversational agent serves as a supportive guide throughout the learning process. Given the current boom in chatbot usage, coupled with the dearth of prior research on chatbots in the field of corporate identity and limited exploration in the realm of adult learning, our study seeks to address the following questions: How do users interact with the assistant and what types of messages are exchanged? The findings of this thesis will shed light on the dynamics of user-agent interaction, the frequency of exchanged messages, and the intended functions of users. As our final product relies on an LLM, which serves as the backbone of the chatbot, we encountered various challenges, such as incorporating external functions and managing the LLM's knowledge limitations. To ensure optimal performance, this thesis includes a comprehensive prompt creation manual, which we utilized to refine our assistant and deliver the most effective learning experience for trainees.

Federated learning has become increasingly more popular due to limitations of the traditional machine learning methods regarding the data privacy. In addition due to technological evolution, the data volume in general has increased by a lot. Mobile devices are capable of storing more and more data. While traditional machine learning methods struggle to deal with these concerns, federated learning emerged from these problems. Two main approaches have been mainly used namely Centralized and Decentralized Federated Learning. The former one has gotten much more attention in comparison with its counterpart and thus possesses many aggregation rules which are resistant to attacks. The goal of this thesis is to propose a new aggregation rule which is resistant to attacks against the machine learning model for the decentralized setting to fill a gap where the research has no reached yet. This is done by extending an existing framework fedstellar, for federated learning. The case studies as part of the evaluation evaluate the algorithm on performance and resource consumption related metrics. They indicate that the performance of the algorithm depends on the situation. They also show the limitation of the algorithm and possibilities of expanding the algorithm to other applications.

Background: Gamification and storytelling have been implemented in slow breathing training apps, a proven effective way to reduce stress and improve mental wellbeing, with the goal of keeping users interested and engaged. However, it is widely acknowledged that users’ cultural background and socioeconomic status (SES) affect their perceptions of apps and therefore intentions to use it on a regular basis. Objective: Based on previous research on user engagement in digital health interventions, cross-cultural human-computer interaction (HCI) and technology acceptance theories, this thesis investigates the intention to use three distinct versions of a slow breathing training app (Stressless©: control group, Breeze©: with gamification, Tragic Kingdom©: with gamification and storytelling) between German and American users. Methods: Adult US and German nationals were recruited from online crowdsourcing platform Prolific. They were randomly assigned to one of the three condition groups and watched a 1-minute introduction video about the corresponding breathing apps in their preferred language. Afterwards, they reported intention to use in a survey, alongside with other attributes related to technology acceptance that are not evaluated in this thesis. As the dependent variable, aggregated intention to use were summed up based on four survey questions. Independent variables were participants’ nationality and app versions coded as categorical variables. Results: A total of 325 participants completed the study (153 German participants and 144 from US Americans). The results show that while national culture does not play a significant role in intention to use, the effect of app version is strong. Although not statistically significant, SES differences in intention to use is observed with no moderating effect of national culture. Conclusion: Even though most hypothesis cannot be rejected; this thesis still provides meaningful guidelines for future studies on cross-cultural and other demographic comparisons on digital health gamification design.

Knowledge workers profit immensely from modern technologies and use them for many aspects of their work. But these technologies also come with drawbacks, as people are now more than ever expected to be available at all times via their work devices. This can lead to blurring boundaries between ones work and private life, which in turn can have negative effects on the ability to psychologically detach from work and enjoy non-work time. In this study we introduce the concept of artefact-based reflection, a method through which knowledge workers can reflect on so called work-artefacts, like tabs, files or e-mails, that they worked with throughout their day and "clean up their workplace" at the end of their workday. Based on existing research and this new concept we developed Downtime, an application which aims at helping knowledge workers in facilitating psychological detachment through artefact-based reflection. We then conducted a user study over two weeks to evaluate the effects of our application. Our findings suggest that knowledge workers can increase their detachment from work and create mental boundaries between their work and non-work lives through reflecting upon the work-artefacts they used that day, especially when they were not actively seeking detachment from their work beforehand. Further research is necessary to evaluate the long-term effects of artefact-based reflection on a broader range of knowledge workers.

Polkadot’s validator selection process employs an iterative algorithm, which is dependent on the size of the staking system. As Polkadot’s staking network is growing, I propose a machine learning alternative approach to the current implementation, that is more independent of scale. The algorithm, the sequential Phragmén, aims to reduce a graph of nominator-validator edges to a subset of validators, the active set, and distribute the stake backing them, as evenly as possible. The goal of this thesis is to produce superior results, consequently improving the overall security or to provide solutions of equal quality in faster time. In order to achieve the goal, a pipeline is setup, that gathers data and transforms it such that it is suitable for machine learning models. Predictions are made, which are adjusted to fit the requirements set by Polkadot. The adjusted results are scored and ultimately compared to the solutions discovered by sequential Phragmén. An analysis of the training data reveals, that the active set remains highly static, with only 10 validators on average changing from era to era. This lack of diversity raised concerns regarding potential attack vectors for adversaries. Furthermore, it was observed that many nominators are acting inefficiently. Many of them do not execute their right to nominate up to 16 validators, which would maximize their chance of having a validator included in the active set. Additionally, many of them include validators, which are not eligible targets. This occurs since nominators frequently ignore their duty to actively tend to their validator preferences. They set them once and do not update them. Eligible validators become inactive (intentionally or unintentionally) and consequently remain as part of the nominators preferences. The prediction task was split up into three models: The first model predicts the next active set, the second model predicts the sum of stake each validator receives and the third predicts the individual stake distribution. The results show, that the first two models are trained well and produce satisfactory results. However, the learning curves of the third model reveal a bias, which make the predictions suboptimal. The source of the bias is likely the substantial changes in target values introduced by a slight shift of active set. We conclude that it is unlikely to outperform the sequential Phragmén using a supervised approach under the described conditions. Therefore, we recommend exploring an unsupervised approach for further research. Furthermore, we recommend the development of a tool for nominators, that could increase the convenience and the security of the overall staking system as a consequence.