Abstract :
[en] Deceptive designs (i. e., dark patterns) are design strategies that manipulate or force users to make decisions against their best interests. Since their impacts caused various harms such as privacy violations, financial loss, etc., and raise legal or ethical concerns, effective dark pattern detection strategies are essential to mitigate the risks they pose. Nevertheless, there are several challenges that hinder effective dark pattern detection, which can be grouped into three problem domains. First, although the diverse nature of dark patterns has been systematically categorized and described in various taxonomies, the assessment of their presence lacks objective criteria and measurement instruments. Second, dataset limitations, such as limited size and diversity, as well as labeling inconsistencies or incompatibilities, create challenges in the development and training
of robust detection tools. They also invalidate a reliable evaluation of the effectiveness of the tools. Third, traditional detection approaches, whether rule-based or single-model AI methods, struggle with the evolving, diverse, and multimodal nature of dark patterns. Additionally, the role of detection tools needs to be clarified, as current solutions offer limited effectiveness, transparency, and explainability. This thesis addresses these challenges through several contributions, adopting an interdisciplinary approach that integrates computer science, artificial intelligence, human-computer interaction, law, and user experience design to ensure both technical rigor and practical usability.
To mitigate subjectivity in dark pattern detection, measurable features are proposed within a systematic and structured framework for one of the most common dark pattern use cases, namely cookie consent processes that allow a more objective assessment. This thesis further examines existing dark pattern datasets and identify critical quality issues, such as limited representativeness and noisy labeling, which have not been systematically presented and analyzed before. In response, a benchmark dataset is built, annotated by dark pattern experts and aligned with a recent unifying taxonomy to mitigate the aforementioned issues. To contribute to the data scarcity problem, a multi-agent framework employing Large Language Models for dark pattern data augmentation is developed, which is validated by showing improved performance of fine-tuned models in the task of dark pattern detection. Finally, the work focuses on Multimodal Large Language Model-based approaches and efficient detection tools are built
by leveraging recent techniques such as Retrieval Augmented Generation (RAG) and Chain-of-Thought (CoT), while integrating the concept of measurable features, which are proposed to address the subjectivity problem, to enhance both accuracy and explainability of the model. The proposed detection strategies were evaluated through quantitative analyses, expert interviews, and empirical comparisons. Furthermore, the study extends to open-source Multimodal Large Language Models, conducting the first empirical evaluations of their performance in the dark pattern detection task, using various prompt engineering strategies and comparing their effectiveness with proprietary models. The thesis concludes with a discussion of the remaining
challenges, limitations, and open problems from an interdisciplinary perspective.
