As smart home environments grow increasingly populated with heterogeneous and resource-constrained IoT devices, ensuring secure and adaptive network protection has become a critical challenge. Traditional intrusion detection solutions, i.e., centralized, rule-based, or cloud-dependent, struggle to scale or provide user-friendly insights for non-technical residents. To address these limitations, we propose HiFINS, a Hierarchical Federated Learning-based Interactive Network Security system that integrates hierarchical federated learning (HiF) for real-time, privacy-aware threat detection and a human-centered interface (INS) for intuitive security management. HiFINS distributes model training across home routers, local servers, and edge/cloud infrastructure, enhancing detection performance while minimizing data exposure. Simultaneously, the interactive application enables users to configure routers, view alerts, and respond to threats without needing technical expertise. The system is implemented and validated within a reproducible, container-based test environment that emulates realistic smart home conditions and facilitates transparent benchmarking. Experimental results show that HiFINS outperforms both centralized and existing hierarchical federated approaches in terms of detection accuracy, inference throughput, and precision, while maintaining competitive training efficiency. These gains demonstrate HiFINS’s practical viability for securing home networks in real time. Although the hierarchical design introduces coordination overhead, the tradeoff enables more resilient and scalable deployment across diverse environments. The system’s modular and reproducible structure supports future adaptation and integration into broader smart home security efforts.

The proliferation of Internet of Things (IoT) devices in smart home environments has significantly expanded the attack surface, making home networks increasingly susceptible to sophisticated cyber threats. Traditional cloud-based intrusion detection systems (IDS) often suffer from high latency and performance bottlenecks, limiting their effectiveness in real-time threat mitigation. This paper introduces a Hierarchical Federated Generative Learning (HFGL) framework for decentralized intrusion detection, leveraging a multi-tiered architecture composed of user routers, a local server, and edge nodes to facilitate real-time network monitoring and distributed anomaly detection. The system integrates OpenWRT-based router firmware with packet-level data capture, securely transmitting network traffic metadata to a local server via SSHbased Paramiko requests. A custom Electron.js-based desktop application provides an intuitive interface for homeowners, enabling seamless router configuration, real-time alert visualization, and security policy orchestration. At its core, our approach employs a federated deep learning pipeline augmented with Generative Adversarial Networks (GANs) to enhance intrusion detection capabilities while preserving data privacy. The GAN model not only identifies emerging threats but also generates adversarial attack simulations to improve model robustness. Experimental evaluations demonstrate significant improvements in detection accuracy and latency reduction compared to conventional IDS approaches, underscoring the potential of privacy-preserving, edge-driven cybersecurity solutions for modern smart home ecosystems.

Effective training in Machine Learning and Deep Learning models necessitates datasets that provide sufficient patterns and contextual information, particularly crucial in IoT networks. Imbalanced datasets, however, significantly challenge the performance of Autonomous Intrusion Detection Systems (IDS), leading to suboptimal detection rates for minority classes. In this paper, we address this issue by utilizing various Generative Adversarial Network (GAN) models, including WGANGP, CGAN, CTGAN, and CWGANGP, to generate synthetic data that balance these imbalanced datasets. We evaluate the performance of IDS models trained on GAN-augmented datasets against those trained on unbalanced datasets, considering metrics such as fitting duration, generation duration, accuracy, precision, recall, and F1 score. Our findings reveal substantial improvements in IDS performance with the application of GANs across binary, general, and specific attack classifications. Additionally, we compare the effectiveness of GANs with classical sampling algorithms, such as SMOTE and Random Oversampling. This comprehensive evaluation underscores the potential of GANs as a sophisticated solution for improving IDS accuracy and reliability in handling complex and highly imbalanced datasets.

Evaluating the relevance of an exogenous data series is the first step in improving the prediction capabilities of a forecast algorithm. Inspired by existing metrics for time series similarity, we introduce a new approach named FARM - Forward Aligned Relevance Metric. Our forward method relies on an angular measure that compares changes in subsequent data points to align time-warped series in an efficient way. The proposed algorithm combines local and global measures to provide a balanced relevance metric. This results in considering also partial, intermediate matches as relevant indicators for exogenous data series significance. As a first validation step, we present the application of our FARM approach to synthetic but representative signals. While demonstrating the improved capabilities with respect to existing approaches, we also discuss existing constraints and limitations of our idea.

Benchmarks for general language understanding have been rapidly developing in recent years of NLP research, particularly because of their utility in choosing strong-performing models for practical downstream applications. While benchmarks have been proposed in the legal language domain, virtually no such benchmarks exist for privacy policies despite their increasing importance in modern digital life. This could be explained by privacy policies falling under the legal language domain, but we find evidence to the contrary that motivates a separate benchmark for privacy policies. Consequently, we propose PrivacyGLUE as the first comprehensive benchmark of relevant and high-quality privacy tasks for measuring general language understanding in the privacy language domain. Furthermore, we release performances from multiple transformer language models and perform model–pair agreement analysis to detect tasks where models benefited from domain specialization. Our findings show the importance of in-domain pretraining for privacy policies. We believe PrivacyGLUE can accelerate NLP research and improve general language understanding for humans and AI algorithms in the privacy language domain, thus supporting the adoption and acceptance rates of solutions based on it.

Everyday we interact with online services from companies that ask for our permission to use our personal information. Nowadays it is common practice for websites and apps to collect big amounts of data which are mainly used for revenue optimization based on user analytics. This customer data collection and usage is regulated by legal agreements (i.e., privacy and cookie policies) which we are required to accept (multiple times a day), but which are generally very long and formulated in a way that makes their interpretation difficult for the general public. An average privacy policy takes 15 min to read and includes lots of legal jargon (e.g., including words like "data controller" and "legal basis for processing"). In this research project, we are developing a support system where users can search for concrete answers in the privacy policies of companies or websites, by formulating their questions in natural language. Instead of blindly accepting a privacy policy, a user could first query the system for answers to a potential concern. The system will return a ranked list of phrases and documents matching the query. In case the generated answer is not sufficient for the user, an extension will allow them to forward complex requests to best-matching legal professionals, specialized in privacy legislation, which can process them for a small fee. We present different aspects of the internal implementation, including the identification of relevant spans in unstructured privacy policies and the selection of the best-suited NLP model for this specific task. The initial results of a user evaluation are presented, showing promising directions. Eventually, some future research directions for the extension of the system conclude our contribution.

Privacy is one of the key issues for citizen's everyday online activities, with the United Nations defining it as "a human right in the digital age". Despite the introduction of data privacy regulations almost everywhere around the globe, the biggest barrier to effectiveness is the customer's capacity to map the privacy statement received with the regulation in force and understand their terms. This study advocates the creation of a convenient and cost-efficient question-answering service for answering customers' queries on data privacy. It proposes a dual step approach, allowing consumers to ask support to a conversational agent boosted by a smart knowledge base, attempting to answer the question using the most appropriate legal document. Being the self-help approach insufficient, our system enacts a second step suggesting a ranked list of legal experts for focused advice. To achieve our objective, we need large enough and specialised dataset and we plan to apply state-of-the-art Natural Language Processing (NLP) techniques in the field of open domain question answering. This paper describes the initial steps and some early results we achieved in this direction and the next steps we propose to develop a one-stop solution for consumers privacy needs.

Similarity metrics measure distance to a compared time series. It allows for a classification and dependency search. These metrics are used for the selection of additional time series in forecasting, which involves advanced information on the target time series, known as exogenous data. Several studies demonstrate an accuracy gain by including such data that represents the environment for the target time series. Yet, robust significance analysis represents a key prerequisite for the correct context identification towards and accurate time series forecast. For this reason, this article presents current similarity metrics and demonstrates significant aspects in aligning time series. The concept of dynamic comparison and its importance will be discussed as a basis for a robust perceptual significance analysis. By employing a pair of exemplary load and exogenous time series, alignment capabilities of promising distance metrics were tested, thus demonstrating gaps for an effective perceptual computing methodology. Finally, this paper examines prerequisites necessary for a new robust significance analysis methodology on three characteristic exogenous and target time-series combinations.

Entity Recognition (ER) within a text is a fundamental exercise in Natural Language Processing, enabling further depending tasks such as Knowledge Extraction, Text Summarisation, or Keyphrase Extraction. An entity consists of single words or of a consecutive sequence of terms, constituting the basic building blocks for communication. Mainstream ER approaches are mainly limited to flat structures, concentrating on the outermost entities while ignoring the inner ones. This paper introduces a partly-layered network architecture that deals with the complexity of overlapping and nested cases. The proposed architecture consists of two parts: (1) a shared Sequence Layer and (2) a stacked component with multiple Tagging Layers. The adoption of such an architecture has the advantage of preventing overfit to a specific word-length, thus maintaining performance for longer entities despite their lower frequency. To verify the proposed architecture's effectiveness, we train and evaluate this architecture to recognise two kinds of entities - Concepts (CR) and Named Entities (NER). Our approach achieves state-of-the-art NER performances, while it outperforms previous CR approaches. Considering these promising results, we see the possibility to evolve the architecture for other cases such as the extraction of events or the detection of argumentative components.

In the context of data access and export control from enterprise information systems, one of the issue is the generation of the rules. Currently, this time consuming and difficult task is highly based on experience. Expert security analysts merge their experience of Enterprise Resource Planning (ERP) systems with the random exploration of the logs generated by the system to try to envision the most relevant attack paths. This project allowed to explore different approaches for creating support for human experts in security rule identification and validation, while preserving interpretability of the results and inspectability of the approach used. This resulted in a tool that complements the security engine by supporting experts in defining uncommon patterns as security-related events to be monitored and vetted by the event classification engine. The result is a promising instrument allowing the human inspection of candidate security-related relevant events/patterns. Main focus being the definition of security rules to be enforced by the specific security engine at run-time. An initial evaluation round shows a positive trend into the users’ perception, even tough a miss of contextual information still hinders its usage by more business-oriented profiles.

Nowadays, it is becoming essential to extract knowledge from diverse, large scale data-sources. An effective approach to make knowledge accessible and providing the necessary means for efficient reasoning to take place is through the use of knowledge graphs. The process of building knowledge graphs is usually focused on generating meaningful representations. Hence, applying structure to it, which takes into account the existence of different knowledge domains, their depth and breadth is mostly disregarded. This particular shortcoming leads to a loss of valuable information that could else be harnessed to provide various additional functionalities to an application. In other words, enhancing knowledge graphs in such a way that they can be explored similar to how Google Maps presents the world to us. By zooming in and out, different relevant aspects become visible while unnecessary noise is blended out. Granular computing by itself is more of a theorem that highlights potential benefits from the application of fuzzy and hierarchical structures. Little is said on how a potential granular knowledge graph can be built and which existing clustering algorithms can be used for this task. As such, this paper aims to provide (1) an in-depth view of which critical requirements need to be met by an algorithm to establish a granular structure, (2) the process for how different commonly used algorithms are coping with them, as well as (3) an overview that outlines the different steps in the process of establishing a granular knowledge structure. Two approaches are identified as the most promising ones: for low dimensional data, a Growing Hierarchical Self-Organizing Map (with its adaptive behaviour) and, in case of data with high dimensionality, one approach from the projective clustering family, thanks to their capability of finding strong correlation in sub-spaces of the original dimensions.

This work focuses on the electric power market, comparing the status quo with the recent trend towards the increase in distributed self-generation capabilities by prosumers. Starting from the existing tension between the intrinsically hierarchical current structure of the electricity distribution network and the substantially distributed and self-organising nature of the self-generation, we explore the limitations imposed by the current conditions. Initially, we introduce a potential multi-layered architecture for a Peer-to-Peer (P2P) energy market, discussing the fundamental aspects of local production and local consumption as part of a microgrid. Secondly, we analyse the consequent changes for the different users' roles, also in connection with some incentive models connected with the decentralisation of the power production. To give a full picture to the reader, we also scrutinise relevant elements of energy trading, such as Smart Contract and grid stability. Thirdly, we present an example of a typical P2P settlement, showcasing the role of all the previously analysed aspects. To conclude, we performed a review of relevant activities in this domain, to showcase where existing projects are going and what are the most important themes covered. Being this a work in progress, many open questions are still on the table and will be addressed in the next stages of the research. Eventually, by providing a reference model as base for further discussions and improvements, we would like to engage ourselves in a dialog with the different users and the broad community, oriented towards a more fair and ecological-friendly solution for the electricity market of the future.

This work focuses on the electric power market, comparing the status quo with the recent trend towards the increase in distributed self-generation capabilities by prosumers. Starting from the existing tension between the intrinsically hierarchical current structure of the electricity distribution network and the substantially distributed and self-organising nature of the self-generation, we explore the limitations imposed by the current conditions. Initially, we introduce a potential multi-layered architecture for a Peer-to-Peer (P2P) energy market, discussing the fundamental aspects of local production and local consumption as part of a microgrid. Secondly, we analyse the consequent changes for the different users' roles, also in connection with some incentive models connected with the decentralisation of the power production. To give a full picture to the reader, we also scrutinise relevant elements of energy trading, such as Smart Contract and grid stability. Thirdly, we present an example of a typical P2P settlement, showcasing the role of all the previously analysed aspects. To conclude, we performed a review of relevant activities in this domain, to showcase where existing projects are going and what are the most important themes covered. Being this a work in progress, many open questions are still on the table and will be addressed in the next stages of the research. Eventually, by providing a reference model as base for further discussions and improvements, we would like to engage ourselves in a dialog with the different users and the broad community, oriented towards a more fair and ecological-friendly solution for the electricity market of the future.

HR tech is a new trend in the hiring process, still facing some inefficiencies and limits. This paper sketches a high level architecture of a trusted support platform for the job placement task. Our design is based on a multifaceted analysis of current practices and requirements from technical, legal, and social perspectives. Relying on the properties of intelligent data analysis approaches, blockchain technology, and distributed identity management, this solution will enable optimisation, compliance, and improvements of the candidate selection process for job openings. The outcomes of the current work will be further applied for the identification of data formats and specific technologies to be used for the implementation. Practical use-cases are currently under development with industry partners.

The rapid and pervasive development of methods from Artificial Intelligence ( AI ) affects our everyday life. Its application improves the users’ experience of many daily tasks. Despite the enhancements provided, such approaches have a substantial limitation in the shortfall of people’s trust connected with their lack of explainability. In natural language understanding ( NLU ) and processing ( NLP ), a fundamental objective is to support human interactions using sense-making of the language for communication. Such methods try to comprehend and reproduce the self-evident processes of human communication. This applies either in receiving speech signals or in extracting relevant information from a text. Furthermore, the pervasiveness of AI methods in the workplace and on the free time demands a sustainable and verified support of users’ trust, as a natural condition for their acceptance. The objective of this work is to introduce a framework for the calculation and selection of understandable text features. Such features can increase the confidence placed into adopted NLP solutions. The following work outlines the Text Feature Framework and its text features, based on statistical information coming from a general text corpus. The showcase experiment uses those features to verify them on the concept recognition task. The results shows their capability to explain a model and its predictions. The resulting concept recognition models are competitive with other methods existing in the literature. It has the definitive advantage of being able to externalize the supporting evidence for a choice of concept identification.

: Electrical power load forecasting defines strategies for utilities, power producers and individuals that participate in a smart grid. While it is well established in planning processes for production and utilities, the importance of accurate forecasting increases for individuals. The ongoing deregulation of the electricity market enables energy trading by individuals, requiring an accurate estimation of the production and consumption. Research on forecast for aggregated demand shows that including features for the forecast from sources, called exogenous, additional to the purely historical consumption data allows to obtain higher accuracy. In fact, their usage demonstrated to be able to explain the large variability observed in the power demand, taking into account the individual influences. Anyway, the influence of exogenous data is hardly investigated for individual forecasting, due to the minor prevalence of this analysis to date. This review shows the benefit of exogenous data usage and the necessity of detailed research on the input features and their influence on detailed, individual level, forecasts of power demand. Eventually, this contribution is concluded by the presentation of open issues and research directions for electric smart communities that the authors would like to address.

An approach to semantic text similarity matching is concept-based characterization of entities and themes that can be automatically extracted from content. This is useful to build an effective recommender system on top of similarity measures and its usage for document retrieval and ranking. In this work, our research goal is to create an expert system for education recommendation, based on skills, capabilities, areas of expertise present in someone’s curriculum vitae and personal preferences. This form of semantic text matching challenge needs to take into account all the personal educational experiences (formal, informal, and on-the-job), but also work-related know-how, to create a concept based profile of the person. This will allow a reasoned matching process from CVs and career vision to descriptions of education programs. Taking inspiration from the explicit semantic analysis (ESA), we developed a domain-specific approach to semantically characterize short texts and to compare their content for semantic similarity. Thanks to an enriching and a filtering process, we transform the general purpose German Wikipedia into a domain specific model for our task. The domain is defined also through a German knowledge base or vocabulary of description for educational experiences and for job offers. Initial testing with a small set of documents demonstrated that our approach covers the main requirements and can match semantically similar text content. This is applied in a use case and lead to the implementation of an education recommender system prototype.

For knowledge management purposes, it would be useful to automatically classify and tag documents based on their content. Keyphrase extraction is one way of achieving this automatically by using statistical or semantic methods. Whereas corpus-index-based keyphrase extraction can extract relevant concepts for documents, the inverse document index grows exponentially with the number of words that candidate concepts can have. Document-based heuristics can solve this issue, but often result in keyphrases that are not concepts. To increase concept precision, or the percentage of extracted keyphrases that represent actual concepts, we contribute a method to filter keyphrases based on a pre–trained convolutional neural network (CNN). We tested CNNs containing vertical and horizontal filters to decide whether an n-gram (i.e, a consecutive sequence of N words) is a concept or not, from a training set with labeled examples. The classification training signal is derived from the Wikipedia corpus, assuming that an n-gram certainly represents a concept if a corresponding Wikipedia page title exists. The CNN input feature is the vector representation of each word, derived from a word embedding model; the output is the probability of an n-gram to represent a concept. Multiple configurations for vertical and horizontal filters are analyzed and optimised through a hyper-parameterization process. The results demonstrated concept precision for extracted keywords of between 60 and 80% on average. Consequently, by applying a CNN-based concept recognition filter, the concept precision of keyphrase extraction was significantly improved. For an optimal parameter configuration with an average of five extracted keyphrases per document, the concept precision could be increased from 0.65 to 0.8, meaning that on average, at least four out of five keyphrases extracted by our algorithm were actual concepts verified by Wikipedia titles.

Document semantic similarity is a current research field, in particular when the concept-based characterization (or signature) of the entities should be automatically extracted from their content. This becomes critical whenever someone would like to build an effective recommender system on top of this similarity measure and its usage for document retrieval and ranking. In this work, our research goal is an expert system for job placement, based on skills, capabilities, areas of expertise present into someone's curriculum vitae and personal preferences. The challenge is to take into account all the personal educational experiences (formal, informal, and on-the-job), but also work-related know-how, to create a concept based profile of the person. This will allow a reasoned matching process with existing job positions, but also towards additional educational experience for profile improvement. Taking inspiration from the explicit semantic analysis (ESA), we developed a domain-specific approach to semantically characterize documents and to compare them for similarity. Thanks to an enriching and a filtering process, we transform the general purpose German Wikipedia dump into a domain specific model for our task. The domain is defined also through a German knowledge base of description for educational experiences and for job offers. Initial testing with a small set of documents demonstrated that our approach covers the main requirements. There are still open issues that we would like to tackle in the next project steps. Alongside, we have other research directions we plan to take into account, ranging from the consideration of information granulation theories to the best parameters set for algorithm tuning, till the extensibility of our solution to a multi-lingual context.

For knowledge management purposes, it would be interesting to classify and tag documents automatically based on their content. Concept extraction is one way of achieving this automatically by using statistical or semantic methods. Whereas index-based keyphrase extraction can extract relevant concepts for documents, the inverse document index grows exponentially with the number of words that candidate concpets can have. To adress this issue, the present work trains convolutional neural networks (CNNs) containing vertical and horizontal filters to learn how to decide whether an N-gram (i.e, a consecutive sequence of N characters or words) is a concept or not, from a training set with labeled examples. The classification training signal is derived from the Wikipedia corpus, knowing that an N-gram certainly represents a concept if a corresponding Wikipedia page title exists. The CNN input feature is the vector representation of each word, derived from a word embedding model; the output is the probability of an N-gram to represent a concept. Multiple configurations for vertical and horizontal filters were analyzed and configured through a hyper-parameterization process. The results demonstrated precision of between 60 and 80 percent on average. This precision decreased drastically as N increased. However, combined with a TF-IDF based relevance ranking, the top five N-gram concepts calculated for Wikipedia articles showed a high precision of 94%, similar to part-of-speech (POS) tagging for concept recognition combined with TF-IDF, but with a much better recall for higher N. CNN seems to prefer longer sequences of N-grams as identified concepts, and can also correctly identify sequences of words normally ignored by other methods. Furthermore, in contrast to POS filtering, the CNN method does not rely on predefined rules, and could thus provide language-independent concept extraction.