DeustoTeka Examinando por Autor "Zulaika Zurimendi, Unai"

Examinando por Autor "Zulaika Zurimendi, Unai"

Mostrando 1 - 12 de 12

Analysing centralities for organisational role inference in online social networks
(Elsevier Ltd, 2021-03) Sánchez Corcuera, Rubén; Bilbao Jayo, Aritz; Zulaika Zurimendi, Unai; Almeida, Aitor
The intensive use of Online Social Networks (OSN) nowadays has made users expose more information without realising it. Malicious users or marketing agencies are now able to infer information that is not published on OSNs by using data from targets friends to use for their benefit. In this paper, the authors present a generalisable method capable of deducing the roles of employees of an organisation using their Twitter relationships and the features of the graph from their organisation. The authors also conduct an extensive analysis of the node centralities to study their roles in the inference of the different classes proposed. Derived from the experiments and the ablation study conducted to the centralities, the authors conclude that the latent features of the graph along with the directed relationships perform better than previously proposed methods when classifying the role of the employees of an organisation. Additionally, to evaluate the method, the authors also contribute with a new dataset consisting of three directed graphs (one for each organisation) representing the relationships between the employees obtained from Twitter.
Analysing the impact of images and text for predicting human creativity through encoders
(Science and Technology Publications, Lda, 2025) Pikatza Huerga, Amaia; Matanzas de Luis, Pablo; Fernandez De Retana Uribe, Miguel; Peña Lasa, Javier; Zulaika Zurimendi, Unai; Almeida, Aitor
This study explores the application of multimodal machine learning techniques to evaluate the originality and complexity of drawings. Traditional approaches in creativity assessment have primarily focused on visual analysis, often neglecting the potential insights derived from accompanying textual descriptions. The research assesses four target features: drawings’ originality, flexibility and elaboration level, and titles’ creativity, all labelled by expert psychologists. The research compares different image encoding and text embeddings to examine the effectiveness and impact of individual and combined modalities. The results indicate that incorporating textual information enhances the predictive accuracy for all features, suggesting that text provides valuable contextual insights that images alone may overlook. This work demonstrates the importance of a multimodal approach in creativity assessment, paving the way for more comprehensive and nuanced evaluations of artistic expression.
Covert L2/L3 tunneling via SIP signaling on embedded hardware: attack, evasion, and detection
(Academic Press, 2026-10) Gorrochategui, Gorka; Zulaika Zurimendi, Unai ; Garaizar, Pablo
Covert channels exploiting application-layer protocols represent a persistent threat in enterprise environments, particularly when targeting endpoints that remain outside the scope of endpoint detection and response solutions, such as embedded Voice over Internet Protocol (VoIP) phones. Existing SIP-based covert channels either embed low-bandwidth steganographic data within call-lifecycle signaling messages (Mazurczyk and Szczypiorski, 2008; Mehić et al., 2014; Tsiatsikas et al., 2015), or achieve higher capacity by requiring active Real-time Transport Protocol (RTP) media streams (Schmidt et al., 2018; Saenger et al., 2020). However, to our knowledge, no prior work implements continuous Layer 2/Layer 3 tunneling over standalone SIP signaling outside of call contexts, nor validates such a channel on real embedded hardware against multiple network intrusion detection systems. We present a covert channel that encapsulates arbitrary Ethernet frames and IP packets within SIP OPTIONS messages, which are lightweight requests typically used for NAT keep-alive and capability probing, without establishing any voice call. A portable proof-of-concept with TUN/TAP virtual interfaces and ChaCha20 symmetric encryption has been implemented and deployed on two ARM-based IP phones from different vendors, showing that even resource-constrained embedded devices can serve as covert network egress nodes. Empirical evasion testing against Suricata, Snort 3, Zeek, and a FortiGate firewall with full Unified Threat Management (UTM) services reveals that the channel achieves a secure operating bandwidth of 128 kbit/s without triggering alerts on any of the tested platforms. Analysis of the tested rulesets and parsers reveals the absence of entropy-based or header-length anomaly rules for Session Initiation Protocol (SIP), accounting for the detection gap. To address this, we implement a lightweight statistical detection mechanism that analyzes non-standard SIP header entropy, showing that the covert channel can be identified through targeted analysis of header content. The gap is not inherent to SIP: current rulesets inspect syntax but not header content, length, or entropy. Adding any of these features is sufficient to detect the channel, as shown by the Zeek script and Suricata 8 rule provided as supplementary material. A single entropy threshold can itself be evaded through low-entropy padding. We therefore show that multi-feature and anomaly-based detectors recover identification of the channel where the scalar test fails. We also contribute a synthetic benign SIP corpus, generated with SIPp to emulate multiple devices and signaling scenarios, released as supplementary material for evaluating false positives beyond the single legacy dataset previously available.
Influence Functions for interpretable link prediction in Knowledge Graphs for Intelligent Environments
(Institute of Electrical and Electronics Engineers Inc., 2022) Zulaika Zurimendi, Unai ; Almeida, Aitor; López de Ipiña González de Artaza, Diego
Knowledge graphs are large, graph-structured databases used in many use-case scenarios such as Intelligent Environments. Many Artificial Intelligent latent feature models are used to infer new facts in Knowledge Graphs. Despite their success, the lack of interpretability remains a challenge to overcome. This paper applies influence functions to obtain the most significant facts when predicting new knowledge and allows users to understand these models. However, Influence Functions do not scale well. We present an efficient method to scale up influence functions to large Knowledge Graphs to overcome such an issue. It drastically reduces the number of training samples when computing influences and uses fast curvature matrix-vector products to linearize the computation steps required for the inverse Hessian. We conduct experiments on different sized Knowledge Graphs demonstrating the scalability of our approach and its effectiveness in measuring the most influential facts. Our method provides an intuitive understanding of link prediction behaviour in Knowledge Graphs and Intelligent Environments.
LWP-WL: link weight prediction based on CNNs and the Weisfeiler–Lehman algorithm
(Elsevier Ltd, 2022-05-01) Zulaika Zurimendi, Unai; Sánchez Corcuera, Rubén; Almeida, Aitor; López de Ipiña González de Artaza, Diego
We present a new technique for link weight prediction, the Link Weight Prediction Weisfeiler–Lehman (LWP-WL) method that learns from graph structure features and link relationship patterns. Inspired by the Weisfeiler–Lehman Neural Machine, LWP-WL extracts an enclosing subgraph for the target link and applies a graph labelling algorithm for weighted graphs to provide an ordered subgraph adjacency matrix into a neural network. The neural network contains a Convolutional Neural Network in the first layer that applies special filters adapted to the input graph representation. An extensive evaluation is provided that demonstrates an improvement over the state-of-the-art methods in several weighted graphs. Furthermore, we conduct an ablation study to show how adding different features to our approach improves our technique's performance. Finally, we also perform a study on the complexity and scalability of our algorithm. Unlike other approaches, LWP-WL does not rely on a specific graph heuristic and can perform well in different kinds of graphs
Machine learning approaches for predicting heart failure readmissions
(Oxford University Press, 2025-07-06) Pikatza Huerga, Amaia; Almeida, Aitor; Quirós López, Raúl; Larrea, Nere; Legarreta Olabarrieta, María José; Zulaika Zurimendi, Unai; García, Rodrigo Damián; García Gutiérrez, Susana
Purpose: This study aims to develop and evaluate machine learning (ML) models to predict the likelihood of hospital readmission within 30 days after discharge for patients with heart failure (HF). The goal is to compare the predictive accuracy of ML models with traditional methods such as those based on Cox proportional hazards and logistic regression, to improve clinical outcomes and reduce hospital costs. Methods: We conducted a prospective cohort study of patients discharged from five hospitals following admission for HF. Data were collected on variables including sociodemographic characteristics, medical history, admission details, patient-reported outcomes, and clinical parameters. ML techniques were employed to analyse the data and predict readmission risk, incorporating strategies to handle class imbalance and missing data. Model performance was assessed based on accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC), and F1 score. Results: Ensemble methods with Synthetic Minority Over-sampling Technique balancing and bagging improved the predictive performance of ML models compared with traditional models. The best-performing ensemble model, using decision trees, Gaussian Naïve Bayes, and neural networks, achieved an AUC of 0.81. In contrast, Cox and logistic regression models showed significantly poorer performance (AUC of 0.58 and 0.50, respectively). SHapley Additive exPlanations analysis revealed that frailty, anxiety, and depression were critical in predicting readmission. Conclusion: ML models, particularly those using ensemble methods, significantly outperform traditional models in predicting short-term readmission for patients with HF. These findings highlight the potential of ML to improve clinical decision-making and resource allocation in HF management.
Machine learning using PROFUND components for 30-day readmission prediction in multimorbid patients: a prospective multicentre study
(Nature Research, 2026) Pikatza Huerga, Amaia; Almeida, Aitor ; Quirós, Raúl ; Legarreta Olabarrieta, María José; Zulaika Zurimendi, Unai ; Mestre, Daniela; García Gutiérrez, Susana
Early hospital readmission in multimorbid patients remains a major clinical challenge. Although risk stratification tools are widely used, predictive performance is often limited. The PROFUND index captures frailty, functional dependence, and social vulnerability, but its role in predicting 30-day readmission is unclear. In this prospective multicentre cohort study, multimorbid patients admitted to Internal Medicine and Geriatrics departments were followed after discharge. The primary outcome was unplanned 30-day readmission among patients surviving to 30 days. Models based on PROFUND components were developed using logistic regression and gradient boosting, including a calibrated ensemble model, and compared with LACE and HOSPITAL scores. Performance was assessed in an external validation cohort. Among 435 patients included in the readmission analysis, 14% were readmitted within 30 days. In external validation, discrimination remained modest (AUC 0.52–0.59). The ensemble XGBoost model achieved the highest AUC (0.59), followed by XGBoost (0.58), HOSPITAL (0.54), and LACE (0.52). Differences were incremental. SHAP analysis identified cognitive impairment, anaemia, advanced age, heart failure severity, functional dependence, and limited caregiver support as key contributors. Incorporating frailty, functional, and social vulnerability domains through PROFUND components resulted in only modest improvements in 30-day readmission prediction. Even with machine learning, discrimination remained limited. The observed performance likely reflects both the intrinsic complexity of short-term readmission and the constraints imposed by sample size and available predictors.
Predictive assessment of eating disorder risk and recovery: uncovering the effectiveness of questionnaires and influencing characteristics
(Elsevier B.V., 2025) Pikatza Huerga, Amaia; Las Hayas Rodríguez, Carlota; Zulaika Zurimendi, Unai; Almeida, Aitor
This study aims to assess the predictive capabilities of various questionnaires in determining the risk of Eating Disorders (ED) and predicting the level of recovery among individuals. Employing machine learning models and diverse datasets, the research focuses on understanding the effectiveness of different questionnaires in providing insights into ED symptoms and recovery outcomes. Additionally, the study seeks to identify the characteristics that significantly influence the recovery process. The investigation aims to contribute valuable information to enhance the diagnostic and monitoring tools used in the field of mental health, particularly concerning ED
Regularized online tensor factorization for sparse knowledge graph embeddings
(Springer Science and Business Media Deutschland GmbH, 2023) Zulaika Zurimendi, Unai; Almeida, Aitor; López de Ipiña González de Artaza, Diego
Knowledge Graphs represent real-world facts and are used in several applications; however, they are often incomplete and have many missing facts. Link prediction is the task of completing these missing facts from existing ones. Embedding models based on Tensor Factorization attain state-of-the-art results in link prediction. Nevertheless, the embeddings they produce can not be easily interpreted. Inspired by previous work on word embeddings, we propose inducing sparsity in the bilinear tensor factorization model, RESCAL, to build interpretable Knowledge Graph embeddings. To overcome the difficulties that stochastic gradient descent has when producing sparse solutions, we add l1 regularization to the learning objective by using the generalized Regularized Dual Averaging online optimization algorithm. The proposed method substantially improves the interpretability of the learned embeddings while maintaining competitive performance in the standard metrics
Smart cities survey: technologies, application domains and challenges for the cities of the future
(SAGE Publications Ltd, 2019-06-10) Sánchez Corcuera, Rubén ; Núñez Marcos, Adrián; Sesma Solance, Jesús; Bilbao Jayo, Aritz ; Mulero, Rubén; Zulaika Zurimendi, Unai ; Azkune Galparsoro, Gorka ; Almeida, Aitor
The introduction of the Information and Communication Technologies throughout the last decades has created a trend of providing daily objects with smartness, aiming to make human life more comfortable. The paradigm of Smart Cities arises as a response to the goal of creating the city of the future, where (1) the well-being and rights of their citizens are guaranteed, (2) industry and (3) urban planning is assessed from an environmental and sustainable viewpoint. Smart Cities still face some challenges in their implementation, but gradually more research projects of Smart Cities are funded and executed. Moreover, cities from all around the globe are implementing Smart City features to improve services or the quality of life of their citizens. Through this article, (1) we go through various definitions of Smart Cities in the literature, (2) we review the technologies and methodologies used nowadays, (3) we summarise the different domains of applications where these technologies and methodologies are applied (e.g. health and education), (4) we show the cities that have integrated the Smart City paradigm in their daily functioning and (5) we provide a review of the open research challenges. Finally, we discuss about the future opportunities for Smart Cities and the issues that must be tackled in order to move towards the cities of the future.
STUN comprehension-optional attributes as a covert channel: design, implementation, and detection
(Elsevier Ltd, 2026-11) Gorrochategui, Gorka ; Zulaika Zurimendi, Unai ; Garaizar, Pablo
Real-time communication protocols carry implicit trust at perimeter defenses because enterprise networks must allow them for Voice over IP (VoIP) and Web Real-Time Communication (WEBRTC) operation. Existing covert-channel literature has explored Session Initiation Protocol (SIP) signaling (Mazurczyk and Szczypiorski, 2008), Real-Time Transport Protocol (RTP)/RTCP media (Mazurczyk, 2013; Schmidt et al. 2018), and WEBRTC streams (Barradas et al. 2020; Figueira et al. 2022), but Session Traversal Utilities for NAT (STUN), despite its central role in NAT traversal and its ubiquity in modern networks, has not been investigated as a covert channel carrier. We design a Layer 2 covert channel that encapsulates arbitrary Ethernet frames within STUN comprehension-optional attributes, which RFC 5389 (Rosenberg et al., 2008) requires compliant implementations to silently ignore. A bidirectional polling mechanism overcomes STUN's client–server asymmetry, producing a full-duplex tunnel indistinguishable from legitimate Interactive Connectivity Establishment (ICE) keep-alive traffic. A portable proof-of-concept was implemented and deployed on two embedded Linux devices of distinct architectures: an ARM-based Yealink T19P E2 IP phone and a MIPS-based Ubiquiti EdgeRouter X. Empirical evasion testing was conducted against Snort 3 with 47,143 Talos rules, Suricata 8 with the ET Open ruleset, Zeek with the spicy_STUN analyzer, and a FortiGate 40F firewall running full Unified Threat Management (UTM) services. Across 3646 data-carrying covert messages, no signature or protocol-aware engine flagged the traffic as anomalous. To address the detection gap, we propose a lightweight Shannon-entropy detector applied to optional-attribute payloads. The mean per-attribute entropy of the covert channel reaches 6.247 bits/byte, well above the highest entropy observed in 1358 legitimate STUN messages from the ITC-Net-blend-60 dataset (4.392 bits/byte from Telegram). A 4.5 bits/byte threshold flags all 3646 covert messages while producing zero false positives across the legitimate corpus. These results expose a systemic detection gap in current network defense strategies and motivate the integration of entropy-aware inspection into Real-Time Communication (RTC) signaling protocols.
Towards more interpretable graphs and Knowledge Graph algorithms
(Universidad de Deusto, 2022-12-13) Zulaika Zurimendi, Unai; López de Ipiña González de Artaza, Diego; Almeida, Aitor; Facultad de Ingeniería; Programa de Doctorado en Ingeniería para la Sociedad de la Información y Desarrollo Sostenible por la Universidad de Deusto
The increase in the amount of data generated by today’s technologies has led to the creation of large graphs and Knowledge Graphs that contain millions of facts about people, things and places in the world. Grounded on those large data stores, many Machine Learning models have been proposed to achieve different tasks, such as predicting new links or weights. Nevertheless, one of the main challenges of those models is their lack of interpretability. Commonly known as “black boxes”, Machine Learning models are usually not understandable to humans. This lack of interpretability becomes even a more severe problem for Knowledge graph-related applications, including healthcare systems, chatbots, or public service management tools where end-users require an understanding of the feedback given by the models. In this thesis, we present methods to increase the interpretability of graphs and Knowledge Graphs based Machine Learning models. We follow a taxonomy grounded on the output result obtained by the proposed methods. Each of the different methods is suitable for particular use cases and scenarios, and can help end-users in different manners. Precisely, we provide an interpretable link weight prediction method based on the Weisfeiler-Lehman graph colouring technique. Additionally, we present an adaption of the Regularized Dual Averaging optimization method for Knowledge Graphs to obtain interpretable representations in link prediction models. Lastly, we introduce the use of Influence Functions for Knowledge Graph link prediction models to acquire the most im- important training facts for a given prediction. Through experiments in link weight prediction and link prediction, we show that our methods can successfully increase the interpretability of the machine learning models of graphs and Knowledge Graphs while maintaining competition with state-of-the-art methods in terms of performance.

Examinando por Autor "Zulaika Zurimendi, Unai"

Resultados por página

Opciones de ordenación