Opin vísindi

 

Nýlega bætt við

Verk
Dual adversarial attacks : Fooling humans and classifiers
(2023-06) Schneider, Johannes; Apruzzese, Giovanni; Department of Computer Science
Adversarial samples mostly aim at fooling machine learning (ML) models. They often involve minor pixel-based perturbations that are imperceptible to human observers. In this work, adversarial samples should fool both humans and ML models, which is important in two-stage decision processes. We perform changes on a higher abstraction level so that a target sample exhibits properties of a desired sample. Technically, we contribute by deriving a regularization scheme for autoencoders incorporating a classifier loss for smoothly interpolating between wildly different samples. The realism and effectiveness of generated samples are confirmed with a user study and other evaluations. Our experiments consider neural networks of four architectures, assessed on MNIST, FashionMNIST, QuickDraw and CIFAR-10. Results show that our scheme leads to superior performance compared to existing interpolation techniques: on average, other methods have an 11% higher failure rate when producing a sample that is of any of two interpolated classes. Furthermore, our attacks work in both white- and black-box settings.
Verk
DReLAB - Deep REinforcement Learning Adversarial Botnet : A benchmark dataset for adversarial attacks against botnet Intrusion Detection Systems
(2021-02) Venturi, Andrea; Apruzzese, Giovanni; Andreolini, Mauro; Colajanni, Michele; Marchetti, Mirco; Department of Computer Science
We present the first dataset that aims to serve as a benchmark to validate the resilience of botnet detectors against adversarial attacks. This dataset includes realistic adversarial samples that are generated by leveraging two widely used Deep Reinforcement Learning (DRL) techniques. These adversarial samples are proved to evade state of the art detectors based on Machine- and Deep-Learning algorithms. The initial corpus of malicious samples consists of network flows belonging to different botnet families presented in three public datasets containing real enterprise network traffic. We use these datasets to devise detectors capable of achieving state-of-the-art performance. We then train two DRL agents, based on Double Deep Q-Network and Deep Sarsa, to generate realistic adversarial samples: the goal is achieving misclassifications by performing small modifications to the initial malicious samples. These alterations involve the features that can be more realistically altered by an expert attacker, and do not compromise the underlying malicious logic of the original samples. Our dataset represents an important contribution to the cybersecurity research community as it is the first including thousands of automatically generated adversarial samples that are able to thwart state of the art classifiers with a high evasion rate. The adversarial samples are grouped by malware variant and provided in a CSV file format. Researchers can validate their defensive proposals by testing their detectors against the adversarial samples of the proposed dataset. Moreover, the analysis of these samples can pave the way to a deeper comprehension of adversarial attacks and to some sort of explainability of machine learning defensive algorithms. They can also support the definition of novel effective defensive techniques.
Verk
Book production in late-fourteenth-century Iceland: A codicological study in the European context
(University of Iceland, School of Humanities, Faculty of Philosophy, History and Archaeology, 2026-07-30) Pokorny, Lea Debora; Beeke Stegmann, Erik Kwakkel; Faculty of Philosophy, History and Archaeology (UI); Deild heimspeki, sagnfræði og fornleifafræði (HÍ); School of Humanities (UI); Hugvísindasvið (HÍ)
The field of Icelandic manuscript studies has become increasingly interested in the material aspects, or codicology, of medieval books, an ongoing trend into which the present study falls. The present thesis aims to contribute to the state of knowledge on vernacular Icelandic manuscript production practice and its organization by centring around the following overarching research questions: (1) How do material characteristics of Icelandic manuscripts dated to the second half of the fourteenth century reflect the methods and practices of their producers? (2) What do observed inconsistencies in codicological features reveal about the resourcefulness, education, and organization of Icelandic book producers? (3) Is it possible to discern to what extent external influences (e.g. other European traditions) shaped production practices of manuscripts in Iceland? At the core of this doctoral thesis stand four peer-reviewed articles and book chapters that address these questions. Two of these were already published by the time of submission, two others have been accepted for publication in peer-reviewed journals. Combined, the four present the data to answer the research questions above. In addition, the introduction to the thesis provides the historical context and methodological framework, while the conclusion puts the findings in the four publications in a broader perspective. To sum up, the results demonstrate that vernacular Icelandic manuscript production during the second half of the fourteenth century utilized a variety of production methods, suggesting that artisans were free to choose which to employ during their work. As such, manuscripts producers in Iceland show inconsistencies in how they proceeded to make handwritten books. The nature of the inconsistencies suggests that the craftsmen aimed to use the resources at their disposal to the fullest, but also that their training was not standardized. Overall, the results speak to a decentralized organization of vernacular book production, which I interpret as “vertical layers”, in which inconsistencies appear to have been the norm, appearing almost naturally as a result of different production practices. While the inconsistencies in production methods can make it challenging to determine the origins of certain practices, there are still clear connections to Insular and the Continental European traditions, demonstrating that Iceland was a sort of “melting pot” of production practices.
Verk
Advancing Clinical Evaluation and Diagnostics with Artificial Intelligence Technologies
(2023-03) Recenti, Marco; Gargiulo, Paolo; Department of Engineering
Machine Learning (ML) is extensively used in diverse healthcare applications to aid physicians in diagnosing and identifying associations, sometimes hidden, between different biomedical parameters. This PhD thesis investigates the interplay of medical images and biosignals to study the mechanisms of aging, knee cartilage degeneration, and Motion Sickness (MS). The first study shows the predictive power of soft tissue radiodensitometric parameters from mid-thigh CT scans. We used data from the AGES-Reykjavik study, correlating soft tissue numerical profiles from 3,000 subjects with cardiac pathophysiologies, hypertension, and diabetes. The results show the role of fat, muscle, and connective tissue in the evaluation of healthy aging. Moreover, we classify patients experiencing gait symptoms, neurological deficits, and a history of stroke in a Korean population, revealing the significant impact of cognitive dual-gait analysis when coupled with single-gait. The second study establishes new paradigms for knee cartilage assessment, correlating 2D and 3D medical image features obtained from CT and MRI scans. In the frame of the EU-project RESTORE we were able to classify degenerative, traumatic, and healthy cartilages based on their bone and cartilage features, as well as we determine the basis for the development of a patient-specific cartilage profile. Finally, in the MS study, based on a virtual reality simulation synchronized with a moving platform and EEG, heart rate, and EMG, we extracted over 3,000 features and analyzed their importance in predicting MS symptoms, concussion in female athletes, and lifestyle influence. The MS features are extracted from the brain, muscle, heart, and from the movement of the center of pressure during the experiment and demonstrate their potential value to advance quantitative evaluation of postural control response. This work demonstrates, through various studies, the importance of ML technologies in improving clinical evaluation and diagnosis contributing to advance our understanding of the mechanisms associated with pathological conditions.
Verk
Advancing Human Soft Tissue Pathology Assessment Using Artificial Intelligence in Medical Imaging
(2024-12) Khatun, Zakia; Gargiulo, Paolo; Department of Engineering
Artificial Intelligence (AI) has revolutionized various fields by automating complex tasks, uncovering patterns in large datasets, and making accurate predictions. Machine learning, particularly deep learning, plays a pivotal role in enabling AI to emulate human intelligence for tasks such as image segmentation, classification, and recognition. In the realm of medical imaging, modalities like Magnetic Resonance Imaging (MRI), Computed Tomography (CT), and Ultrasound are indispensable tools for diagnosing pathologies, detecting abnormalities, guiding treatment plans, and monitoring disease progression. The integration of AI with medical imaging offers the potential to enhance the accuracy and efficiency of these processes. In particular, AI’s application in the assessment of tendon-related conditions, such as tendinopathy, presents a promising avenue for improving patient care. Tendinopathy can significantly impact a patient’s quality of life, and early detection is crucial to optimize treatment outcomes. This thesis focuses on the development of advanced AI-driven methods for analyzing human soft tissue pathologies, with a primary emphasis on tendon segmentation, pathology detection (classification), and tendon reflex response assessment. By automating the analysis of tendons and other human soft tissues, these methods aim to reduce human error and variability, thereby enabling more consistent and reliable clinical decisions. Ultimately, the goal is to support earlier, more accurate diagnoses and interventions, leading to better patient outcomes and more personalized treatment strategies. This thesis begins with a study that analyzes MRI and CT scans from 47 participants to investigate the relationships between the tendons, cartilage, and muscles in the knee. This study has two primary objectives: first, to predict knee cartilage degeneration, and second, to predict patellar tendinopathy. For both objectives, predictions are made using features extracted solely from the patellar tendon and quadriceps, rather than directly from the cartilage itself. This approach explores the potential of using features from surrounding tissues as indirect predictors of knee-related pathologies. This study demonstrates that both knee cartilage degeneration and patellar tendinopathy can be predicted using these features from adjacent structures, highlighting the importance of surrounding tissues as potential indicators of pathology. Traditional machine learning models are employed to identify the most relevant features for each prediction task, highlighting their importance in the diagnosis of these conditions. This foundational research deepens our understanding of the interrelationships between knee soft tissues, contributing to more accurate diagnostic approaches in musculoskeletal health and enhancing clinical decision-making and treatment strategies. A central focus of this thesis is the development of an end-to-end tendon segmentation module. This system integrates a superpixel-based coarse segmentation step that serves as a foundation for the final, more precise segmentation. In this approach, the segmentation task is framed as a superpixel classification problem. To achieve this, two distinct approaches are developed: (1) Random Forest (RF) and Support Vector Machine (SVM) classifiers for superpixel categorization, and (2) a Graph Convolutional Network (GCN) for transforming superpixels into graph structures for node classification. The RF and SVM classifiers demonstrate exceptional performance, achieving Area Under the Curve (AUC) scores of 0.992 and 0.987, respectively, with high sensitivity, indicating their effectiveness in accurately classifying superpixels. Although the GCN approach yields slightly lower performance, it showcases the potential of deep learning methods for improving segmentation by leveraging the structural relationships between superpixels. The findings suggest that both traditional machine learning and deep learning techniques offer promising avenues for advancing tendon segmentation, with superpixel-based methods offering a pathway to more reliable and automated segmentation in medical imaging. Another key component of this thesis is the development of an end-to-end tendon pathology detection module, utilizing the same MRI dataset. This module adopts a graph-based approach, where superpixels are treated as nodes and connected by edge relationships. Each MRI scan is transformed into a graph, with the task framed as a graph classification problem to determine the presence or absence of pathology. To achieve this, a Graph Echo State Network (GESN) is employed. Known for its ability to efficiently represent data without the need for iterative backpropagation, the GESN leverages both temporal and structural dependencies in the data, enhancing classification performance. In this study, the GESN outperforms traditional machine learning models, achieving a mean accuracy of 0.953 and sensitivity of 0.943. These results underscore the potential of the GESN to significantly enhance diagnostic accuracy, offering a powerful tool for early detection and clinical decision-making in tendon pathology assessment. Moreover, the GESN’s ability to handle complex, high-dimensional data suggests its broad applicability to other medical imaging tasks, further expanding its potential clinical utility. The final study of this thesis explores the impact of demographic factors, including age, height, weight, and gender, on reflex response times in healthy individuals. This analysis is based on electromyography (EMG) recordings from 40 participants. The results reveal that elderly individuals, particularly those who are taller, heavier, and male, exhibit delayed reflex onsets. Even after normalizing for height, older participants still demonstrate slower reflex responses. These findings highlight the role of demographic factors in neuromuscular reflexes, aiding in the diagnosis and early detection of related disorders. In conclusion, this research demonstrates the potential of AI, particularly superpixel-based and graph-based models, to advance tendon pathology assessment and exploratory tendon reflex studies, leading to better patient outcomes and musculoskeletal health management.

Flokkar í Opnum vísindum

Veldu flokk til að skoða.

Niðurstöður 1 - 9 af 9

Öll gögn í Opnum vísindum eru vernduð af ákvæðum höfundalaga og með öllum réttindum áskildum, nema annað sé tekið fram.