Large Language Models (LLMs) such as GPT-4 and LLaMA have demonstrated promising performance in fact-checking tasks, particularly in labeling the veracity of claims. However, the real-world utility of such fact-checking systems depends not only on label accuracy but also on the faithfulness of the justifications they provide. Prior work has explored various prompting strategies to elicit reasoning from LLMs, but most studies evaluate these styles in isolation or focus solely on veracity classification, neglecting the impact on explanation quality. This study addresses that gap by investigating how different prompt styles affect both the accuracy and the faithfulness of LLM-generated claim labelling and justifications. Seven established prompting strategies such as Chain-of-Thought, Role-Based, or Decompose-and-Verify, were tested across two datasets (QuanTemp and HoVer) using two efficient models: LLaMA 3.1:8B and GPT-4o-mini. Additionally, two novel prompt variants were introduced and all styles were tested under three label conditions to assess bias and explanation drift.

Large Language Models (LLMs) are increasingly used to generate fact-checking explanations, but evaluating how faithful these justifications are remains a major challenge. In this paper, we examine how well four popular automatic metrics—G-Eval, UniEval, FactCC, and QAGs—capture faithfulness compared to expert-written explanations. We look at how these metrics agree with each other, how they correlate with explanation similarity, and how they respond to controlled errors.Our findings show that while some metrics like UniEval and FactCC show some sensitivity to noise and partial alignment with expert reasoning, none of them reliably catch hallucinations or consistently reflect true faithfulness. Manual analysis also reveals that metric behavior varies depending on the type and structure of the claim. Overall, current metrics are only moderately effective and often biased toward the style of LLM-generated text. This study points to the need for more reliable, context-aware evaluation methods and offers practical insights for improving how we measure faithfulness in fact-checking tasks.

Large Language Models (LLMs) are becoming more commonplace in today's society. However their adoption rate, especially in the fact checking field, is being slowed down by the distrust in their thinking process and the rationales leading to the results. In crucial moments the justifications behind a verdict are more important than the verdict itself. However, LLMs often produce explanations that are not grounded in the provided evidence, leading to hallucinations and reduced trust in their outputs. This paper aims to show exactly the level the LLMs have reached in both the faithfulness of their explanations, based on some provided facts, and the correctness of their explanations. To investigate this, multiple LLMs are asked to assign a label to a claim based on some evidence provided from two datasets of varying complexity: HoVer and QuanTemp. The outputs are then evaluated both manually and by another LLM to evaluate how well the LLM relates to the evidence and if the LLM hallucinates in some parts of its responses. The results reveal that while some models demonstrate high correctness in label assignment, faithfulness in explanations varies significantly across models and evidence types. The outcomes of this experiment aim to inform both LLM developers and fact-checking researchers about the current limitations of LLMs in response quality while also showing which areas require further improvements to become mainstream.

Secure Multi-Party Computation (SMPC) is a widely-used cryptographic tool for privacy-preserving data analysis. The progress in the field of quantum computing has led to the development of Quantum SMPC (QSMPC), which promises informationtheoretic security based on physics laws. There is a significant number of proposed protocols. However, as the number of theoretical protocols grows, their practical viability remains unclear. Our paper presents a systematic literature review of 37 recent QSMPC protocols to assess the state of the field, focusing on protocol functionality, quantum resource requirements, privacy-ensuring techniques, and feasibility on current technologies. Our analysis reveals a plethora of innovative theoretically-private protocols, constrained by significant practical issues. We found that protocols tend to rely on a core set of quantum resources, including entangled states (Bell, GHZ), Quantum Key Distribution, and decoy particles to ensure security. We also observe a frequent reliance on a semi-honest third party. Most importantly, we identify a noticeable gap in feasibility. 19 out of 37 of the papers surveyed provide no clear discussion on the practicality of the proposed protocols. The papers that do discuss this topic describe small-scale simulations, which may require algorithmic compromises and are tested in ideal conditions only. Our findings highlight a gap between theory and practice, suggesting that real-world application of the protocols is not yet possible, given the current state of quantum hardware.

Homomorphic Encryption (HE) enables computation directly on encrypted data, while offering strong cryptographic and privacy guarantees for data-driven sectors like healthcare, finance, and cloud computing. However, practical adoption of HE is severely limited by its computational overhead and specialized expertise requirements. This thesis investigates the financial feasibility of HE by analyzing its Total Cost of Ownership (TCO), an overlooked but crucial factor for researchers, companies, and cloud service customers deciding whether to adopt HE. Building on performance benchmarks from Guillot et al. and Lo et al., we evaluate the runtime, resource requirements, and personnel costs of Fully Homomorphic Encryption (FHE) across different workloads, and Amazon Web Service (AWS) cloud configurations. Our results show that encryption and expert labor dominate the cost, with encryption accounting for over 80% of the runtime and personnel expertise exceeding 98% of the total annual cost ($125,558.066). These findings highlight that, despite the security benefits of HE, its financial overhead restricts its adoption to organizations with substantial cryptographic expertise and budgets. This thesis equips researchers and companies with data-driven insight into HE's true economic impact and underlines the need for further optimization and automation to make HE accessible and cost-effective for broader, real-world use.

Differential Privacy (DP) has become one of the most used approaches to protect individual data. However, its implementation can vary significantly depending on the context we are using it. In this study, we aim to compare two such implementations of DP: Google's Differential Privacy, an open-source library used for structured data analytics, and Differentially Private Offsite Prompt Tuning (DP-OPT), a tool used for the adaptation of machine learning models.

The main research question of this study is the following: "How do DP-OPT and Google's Differential Privacy Library compare when accounting for different factors in different contexts?" We aim to conduct this research by doing a literature-based study where no empirical experiments will be performed. This is because of the underlying complexity of both tools, especially DP-OPT, and the time constraints posed in this project.

The main results of this study show that Google DP is a tool that benefits from its interpretability and speed when processing data analysis, while DP-OPT shows a higher accuracy when training large language models (LLMs). Meaning that there is no single mechanism for each problem, but instead it depends on factors like the underlying task complexity, current available resources, and the final goal of the task.

By comparing these tools side-by-side, this research aims to provide more insights into how each tool behaves and performs under different contexts and tasks. We aim to guide developers and researchers to make better and more informed decisions when choosing a tool for their desired task.

Secure multi-party computation (SMPC) is a cryptographic technique that enables multiple parties to work together on data without sharing their private information with each other. This paper investigates how two open-source frameworks, SecretFlow and FATE, implement SMPC and other privacy-preserving techniques and evaluates their usability and scalability in applied settings. The goal is to better understand the trade-offs these frameworks offer for researchers and developers seeking to build privacy-preserving machine learning pipelines. Using a small-scale experimental setup, both frameworks were deployed in controlled environments and evaluated based on setup complexity, documentation quality, modularity, architecture, and the secure computation methods used. The results show that SecretFlow offers greater SMPC flexibility, modularity, and developer usability, making it well-suited for research and rapid prototyping. FATE, while more complex to integrate, provides comprehensive workflow coordination and is better suited for production environments. Experiments were conducted on a single physical node with a small dataset, enabling reproducibility but limiting generalizability to large-scale, real-world deployments. This work provides practical insights into the usability and architectural design of the frameworks, along with a protocol-level comparison to guide developers in selecting tools for privacy-sensitive machine learning.

This thesis presents a computational framework aimed at enabling the analysis and modeling of boomerangs from example shapes. The goal is to provide a systematic and data-driven tool for boomerang design based on real-world geometries. A key challenge in this context is establishing accurate shape correspondences between handcrafted, asymmetric, and aerodynamically functional boomerangs.
To address this, the proposed pipeline integrates multiple components: landmark-based pre-alignment, boundary extraction using alpha shapes, curve parameterization, and Least Squares Conformal Mapping (LSCM) to compute surface correspondences. Building on these correspondences, the framework further incorporates principal component analysis (PCA) and Free-Form Deformation (FFD) to enable the generation of new shapes.
Experimental results show that the method achieves low Hausdorff and Chamfer distances and has been evaluated using area- and shear-based distortion metrics. Nonetheless, some localized inaccuracies - particularly near high-curvature regions - highlight areas for improvement in boundary handling and local control. Additional limitations include the reliance on manual landmark selection, the linearity of PCA, and the sensitivity of the pipeline to alpha shape parameters.
By combining elements of computational geometry with principles of aerodynamic design, this research bridges the gap between empirical craftsmanship and formal shape analysis. The resulting methodology not only enables the comparison and modeling of boomerangs but also lays the groundwork for future tools in boomerang shape exploration and design.

Alzheimer’s disease (AD) is a neurodegenerative disorder prevalent in older adults, leading to loss in memory, cognitive, and executive function. A characteristic feature of AD is the accumulation of amyloid-beta (Aβ) plaques, which are extracellular deposits of Aβ protein primarily found in grey matter. These Aβ deposits can appear in different forms. The six primary Aβ deposits covered in this work are: diffuse plaques, cored plaques, compact plaques, coarse grained plaques, cerebral amyloid angiopathy (CAA), and subpial deposits. It is speculated that some of the plaques types may be more harmful than others. Given that AD primarily affects an older population, the question arises of how individuals with AD differentiate from cognitively healthy centenarians (people over 100 years old). Consequently, this work attempts to classify and analyse different Aβ types present in donated brain tissue of cognitively healthy centenarians who escaped dementia and individuals diagnosed with AD. The main goal is to identify differences between these two cohorts. However, a challenge in this process is that the brain tissues contain many Aβ plaques, making manual identification difficult in terms of time and labor. To address this, a fine-tuned ResNet50 Aβ plaque classifier was developed in this research that was integrated into an Aβ detection pipeline capable of localising plaques. The model was initially pre-trained on the ImageNet dataset through contrastive learning, and subsequently fine-tuned using few-shot learning with a small number of annotated samples (315). The annotations included the six primary Aβ types whose structure are known and well-defined, and three other anomaly Aβ types that
served to filter out irregular plaques in the unlabeled Aβ data. After performing 5-fold cross-validation, the fine-tuned models demonstrated an average accuracy of 85.71% and precision of 89.47% on the primary types. The final classifier used on the unlabeled Aβ dataset was an ensemble model that incorporated majority voting, combining the predictions of the five models trained during cross-validation. Aβ loads were calculated for each primary Aβ type based on the classifier’s predictions. It was observed that across all considered primary Aβ types, the centenarians’ Aβ loads were statistically significantly lower compared to the AD cohort. The lower Aβ load in centenarians also held true across the frontal, temporal, parietal, and occipital cerebral regions for each primary Aβ type. To partially validate the
model’s performance, correlations were computed between the predicted Aβ loads of the primary Aβ types and neuropathological assessments collected from 75 centenarians. These assessments are common in related works and serve as a benchmark for the ensemble model. They include the Thal Aβ phase, which categorizes the distribution of general Aβ in the brain; the Thal CAA stage, which measures the severity of CAA; and CERAD NP scores, which evaluates the spread of neuritic plaques in the brain, which are a subset of cored plaques. Consequently, statistically significant positive correlations were revealed between: the Thal Aβ phase and the Aβ load of all primary types (r ranging 0.59-0.73); the Thal CAA stage and Aβ load of predicted CAA deposits (r = 0.66); and the CERAD NP scores and Aβ load of cored plaques (r = 0.68). Since the correlated types coincide with what the benchmark staging schemes measure, the model’s predictions seems to align with existing literature.

An accurate segmentation model for hip compo- nents could improve the diagnosis of Osteoarthritis, a prevalent age-related condition affecting joints. A significant challenge in developing effective and robust segmentation models are the domain differ- ences across various datasets. In this study, we in- vestigate the impact of different data augmentation and preprocessing techniques on the generalizabil- ity of femur segmentation models across datasets. Using two labeled datasets, we evaluate the perfor- mance of a U-Net segmentation model, focusing on the effectiveness of augmentations like image flip- ping, random rotations, blur, contrast, and bright- ness adjustments. Our findings reveal that certain augmentations, particularly random rotations of up to 15 degrees, vertical image flipping and light blurring, significantly improve the model’s gener- alization to another data set, reducing boundary er- rors and enhancing segmentation accuracy. These results underscore the potential of targeted data augmentations in developing robust, generalizable models for hip joint component segmentation.

The severity of hip osteoarthritis is measured a.o. by the minimal distance between the femoral head and the acetabular roof in an X-ray image. However, the whole joint space profile might be a more accurate estimator, since it would include irregularities in the bone surface. These irregular bulges (osteophytes) on the bone surface are one of the signals that a person might have OA. Thus the stage of OA might be better estimated automatically by having this data in the joint space profile instead of just using the minimal joint space.

For this joint space profile, the distance between the femoral head and the acetabular roof needs to be calculated. Therefore, the positions of these parts in the hip joint are required to be know. These can be retrieved from e.g. a segmentation mask.

One way of calculating the distance in a joint is to use a radial projection. A radial projection is a way of projecting points from a curved space to a plane by projecting lines from a central point along increasing angles.

In this paper, we investigate how the joint space profile can be segmented most accurately from a radial projection originating from the center of the femoral head by several comparing noise filtering and edge-finding algorithms. After which is shown that a custom algorithm based on the theory behind edge detection in noisy images works most reliably and accurately.

There are still multiple points of improvement for this algorithm. The femoral head can be segmented more accurately than the acetabular roof, the segmentation of the latter could be optimized by detecting the brightest line (peaks) instead of the most sudden change (steepest gradient) in the X-ray image as the edge for the femoral head. The algorithm could be further improved by taking care of local outliers off those edges.

In conclusion, this paper compares multiple ways of segmenting the joint space of the hip joint. The best-performing algorithm could in the future be used in an assisting tool for doctors to highlight important irregularities and measurements in the hip joint space.

Osteoarthritis is a degenerative disease that affects the aging population by degrading the cartilage in the joints. The early and accurate diagnosis of this disease is key to effective treatment. For an early and accurate diagnosis of this disease, clinicians often use X-ray imaging. This allows medical professionals to manually measure the joint space width (JSW) in X-rays images to determine the progression of the disease. This method however proves to be both time-consuming and variable based on the professional. This research addresses the automation of the measurement of the JSW for the hip, using deep learning techniques, to improve precision and efficiency. The automated measurement of the JSW is challenged by variations in the imaging conditions across different clinical settings. To address these discrepancies and keep a good performance, domain adaptation techniques are used to counter these domain shifts to ensure a consistent JSW segmentation across different imaging domains. The study investigates whether a specific domain adaptation technique can enhance the accuracy and robustness of deep learning models specifically for femur segmentation in X-ray images across different datasets. A base deep learning model is developed for femur segmentation, and supervised domain adaptation is applied. The study compares the performance of the adapted model with the base model across two different datasets. Results indicate that supervised domain adaptation does not significantly improve the model’s robustness and accuracy in femur segmentation among two different datasets. These unexpected findings suggest that incorporating domain adaptation techniques may not always lead to a more reliable and efficient diagnosis of osteoarthritis, reducing the manual workload for clinicians.

Deep learning based architectures have been applied to semantic segmentation tasks in medicalimaging with great success. However, such modelsare heavily reliant on the quality of the groundtruth segmentation mask and hence are susceptibleto label noise. To address this issue, thispaper introduces SuperLoss, a loss function thatpushes semantic boundaries towards superpixeledges. Superpixels are compact, homogeneous regionswithin an image that group pixels with similarcharacteristics, such as pixel intensity. Our losscan be combined with other loss functions for differentsegmentation architectures. We demonstrateour framework on a combination of two large publicdatasets of hip joint X-Ray images. We comparea U-Net model with and without our loss,when trained with different fractions of noise in thetraining dataset. Our approach achieves a 1 − 2%improvement in Intersection-over-Union and Hausdorffdistance for some cases, yet yields worse insome other cases. We also perform hypothesis testingand show that our results are statistically significantwith low to medium effect size.

Hip osteoarthritis is a widespread disease, with medical experts facing difficulties in this illness, due to a lack of standard grading score. Nevertheless, the minimum joint space width remains the most important score for osteoarthritis severity. Manual estimation of this metric is a tedious task, which can greatly benefit from employing an automated tool. While some research has been done in developing such a tool using deep learning methods, a novel and promising approach, most lack annotated data to use for training, which can be hard to obtain. Thus, thus research aims at developing a deep learning approach towards estimating the minimum joint space, while using automatically labels produced with an existing algorithm.

This study addresses the gap for fine-grained emotion recognition in immersive environments utilizing solely data from on-board sensors. Two data representations of users eyes are utilized, including periocular recordings and eye movements (gaze estimation and pupil measurements). A novel multi-representation method integrating feature extractors for each representation alongside an effective feature fusion technique is proposed. The method significantly outperforms baselines that use only a single representation or incorporate content stimuli. It achieves an F1-score of 0.85 with 10\% data, approximately 40 seconds of data from all emotions, for personal adaptation, recognizing emotions while watching unseen parts of the stimuli used for adaptation. In a more practical scenario, the method achieves an F1-score of 0.71 with five seconds of personal adaptation data from each emotion, recognizing emotions while watching completely unseen stimuli. Under the same but more extreme condition, where only one second of data is available, the proposed achieves an F1-score of 0.68.
Furthermore, the study demonstrates that estimated labels can substitute for user-provided labels without sacrificing performance in emotion recognition, thus eliminating the need for users to manually label emotion elicitation segments.
Future work will focus on improving performance by allocating more computational resources and making architectural modifications, conducting deeper investigations into the decision-making process, and developing real-time recognition systems for in-the-wild experiments.
The results of this study suggest that more engaging, adaptive, and personalized experiences in immersive environments can be developed.

Alzheimer's disease (AD) is becoming more prevalent as the world population gets older. The formation of Amyloid-beta (\AB) plaques is one of the pathologies related to AD. Recent work has shown that the \ab load in brain tissue has a negative correlation with cognitive performance in cognitively healthy centenarians.
This work aims to expand this research by investigating whether the types of \ab plaque present are linked to cognition and by comparing the types of plaques in the centenarian cohort with an AD cohort.
For this task, a system is developed that can identify \ab plaques in images of brain tissue. It first automatically segments the grey matter using a fine-tuned U-net. Then the plaques are located using traditional image processing techniques. Lastly, shape and size features are extracted from the plaque in addition to a feature vector made by a pre-trained AlexNet. K-means clustering is used on AlexNet features to find categories for the plaques.
The clustering approach failed to yield good results. However, the area and roundness are differently distributed between the AD and centenarian cohorts, but the differences are small. Correlations have been found between the area and roundness of plaques in the occipital pole and cognitive performance in centenarians. They indicate that cognitively stronger individuals have smaller and less round \ab plaques in their brains. More research is necessary to reveal the true extent of the impact of plaque types on cognition.