Towards Trustworthy Artificial Intelligence - Insights from Selected Studies

The development of AI has experienced a significant acceleration over the past decade. In recent years, this trend has been particularly reinforced by the rise of generative AI. As a result, there has been a notable improvement in the performance of AI systems, which have been deployed in a variety of application areas. For instance, in the field of marketing, generative AI is utilized to create personalized content and advertisements. In industrial settings, predictive maintenance employs AI to assist companies in anticipating equipment failures. Nevertheless, numerous instances have illustrated that AI systems can potentially pose significant risks. For example, commercial facial recognition systems have been demonstrated to exhibit substantial racial bias, with individuals with darker skin tones being misidentified at significantly higher rates than those with lighter skin tones. Furthermore, AI systems have exhibited a lack of interpretability, which presents a significant risk in applications where users require an understanding of the AI's decision-making process. For instance, in healthcare, a doctor must be able to comprehend why an AI-based diagnostic tool might provide a specific treatment recommendation, as the decision can have a profound impact on a patient's life. In academia, several research fields aim to better understand the risks of AI and develop strategies to mitigate them. Notable fields include explainable AI, which explores how to make AI systems more understandable; privacy-preserving machine learning, which concentrates on protecting privacy within AI systems; and fair machine learning, which seeks to prevent AI from exhibiting biased behavior towards certain user groups. Significant progress has been made in these fields over the past years. Recently, additional research efforts have emerged that adopt more comprehensive approaches, extending the scope of existing research with new perspectives. Many of these activities have been addressed within the context of Trustworthy AI. This emerging field of research is characterized by a holistic perspective, examining a wide range of aspects related to the development and operation of trustworthy AI systems. For example, it investigates the concept of trust in the context of AI, considers the criteria by which an AI system can be deemed trustworthy, and examines the challenges in developing trustworthy real-world applications. This dissertation aims to contribute to the field of Trustworthy AI by presenting an overview of its current state of research and providing insights into selected issues. It describes the results of five largely independent studies that address specific issues related to Trustworthy AI. Study 1 aims to provide an overview of the current state of research within the field of Trustworthy AI. First, a meta-review is conducted to give an overview of the research landscape and systematically highlight the similarities and differences among existing reviews. The results demonstrate that while several areas of Trustworthy AI have been commonly addressed, there is a limited understanding of the feasibility of combining multiple aspects of trustworthiness, potential synergies or trade-offs between them, and the technical challenges involved. To gain a deeper understanding of this issue, an additional review is conducted to explore how multiple aspects of trustworthiness are currently integrated. The results reveal a growing interest in comprehensive approaches to Trustworthy AI; however, their implementation remains a significant challenge. Study 2 investigates the evaluation of Trustworthy AI, which can help to get an impression of the trustworthiness of AI systems before they are deployed. Initial approaches for evaluating Trustworthy AI are characterized by the fact that, in addition to model performance, they assess various aspects of trustworthiness in AI systems. A perspective that has received limited attention is the potential for quantitative approaches that could allow better operationalization of trustworthiness. While previous quantitative approaches focused on individual aspects of trustworthiness, initial studies have started to investigate how to quantify multiple aspects simultaneously. This study aims to investigate the feasibility of quantitatively evaluating multiple aspects of trustworthiness together and whether it can provide an overall assessment of an AI system's trustworthiness. In an experimental approach, several image classifiers are trained in a practical context. To evaluate the models' trustworthiness, a novel framework is developed that quantitatively assesses the aspects of explainability, fairness, privacy, and robustness in a balanced way and aggregates them into a comprehensive score. The findings indicate the feasibility of the developed framework to provide an overall assessment of trustworthiness and its likely transferability within image classification. Study 3 builds upon Study 2 and explores the interdependencies between the four evaluated aspects of trustworthiness. In practice, the evaluation framework of Study 2 might show that one particular aspect of trustworthiness does not meet the desired standards. In order to counteract this, several modifications to the data set, model, or training process have been examined. While existing literature has largely focused on the effectiveness of these interventions, one perspective that has been explored less is the extent to which improving one aspect of trustworthiness might affect others. Prior research has acknowledged that aspects of trustworthiness are not independent; however, comprehensive studies examining their interdependencies are scarce. This study investigates how modifying training processes to enhance one aspect of trustworthiness impacts others. In an extensive set of experiments, several image classifiers are trained while enhancing one trustworthiness aspect at a time. The models are quantitatively evaluated using the framework developed in Study 2. The study uncovers several interdependencies and trade-offs between explainability, fairness, privacy, and robustness. The results suggest that the chosen approach is suitable for quantifying these effects and that such investigations could be valuable in application scenarios where multiple trustworthiness aspects are essential. Study 4 addresses the issue of distribution shifts, where model performance degrades as operational data diverges from training data. One approach to handle distribution shifts is to make models more robust, thereby prolonging their operational lifetime. Recent research has used model explanations not only to understand models better but also to leverage them for adjustments that enhance performance. Study 4 aims to investigate the feasibility of using explainability to improve robustness against distribution shifts. Hyperparameter tuning is known to affect robustness due to its impact on model generalization. In this study, explanations are used to guide the tuning process by providing new insights into model behavior to improve robustness. The effectiveness of this approach is demonstrated on three datasets from different contexts. Study 5 examines the challenges of Trustworthy AI in the rapidly growing domain of AI-as-a-Service (AIaaS), which represents an AI-specific counterpart to the prominent software-as-a-service model. While AIaaS offers companies easy access to advanced AI capabilities without relying on large amounts of skill or capital, there could be challenges to trustworthiness due to how these systems are operated. For instance, developers lack control over the circumstances under which their models will be used. This study aims to shed light on these considerations and explore the unique challenges associated with Trustworthy AI in the AIaaS domain. The study begins with a systematic literature review, identifying that while the concept of Trustworthy AI has been sparsely studied in the context of AIaaS, several challenges related to trustworthiness are already emerging in the literature. The review is followed by an experimental investigation of an AIaaS solution for mobile device identification, finding several implications related to Trustworthy AI, for example, with regard to accountability, privacy, or the need for measures against potential misuse.