Abstract Named Entity Recognition (NER) in specialized domains like biomedicine suffers from acute data scarcity, requiring expensive expert annotations. While data augmentation offers a promising solution, it inevitably introduces noisy and mislabeled samples that can degrade model performance. This problem is amplified in few-shot scenarios where every training example matters. We introduce PALAUNER (Policy-based Active Learning to Augment Named Entity Recognition), a reinforcement learning framework that learns to select high-quality samples from augmented data pools. Using a deep Q-network, our agent evaluates samples based on content features and model predictions, deciding which examples will improve NER performance. Experiments across five BioNER benchmarks demonstrate that PALAUNER consistently enhances diverse augmentation methods, from simple perturbations to GPT-based generation. Average F1 improvements are of 0.5−7.1 points in few-shot settings. PALAUNER’s modular design enables seamless integration with emerging augmentation techniques, providing a generalizable solution for training data quality enhancement. We publicly release our code on GitHub: ( https://github.com/picuslab/palauner ).

Financial Natural Language Processing crucially leverages social media for market insights. However, most existing methods for this purpose rely on simple sentiment analysis models, which fail to capture the concrete trading intentions expressed in these discussions. While Large Language Models (LLMs) offer a promising alternative to simplistic sentiment analysis, the actual benefits of fine-tuning across different model families remain unclear in noisy, domain-specific contexts like online forums. To address this gap, we present a comprehensive assessment of the advantages and limitations of fine-tuning for Social Trading Action Detection (STAD), a novel task that aims to classify online posts into actionable categories, namely buy, sell, or other. In addition, we introduce FinReddit-2K, a manually annotated dataset consisting of 2123 Reddit posts, designed to serve as a benchmark for this task. Our experimental analysis goes beyond standard performance metrics and identifies both the types of errors that fine-tuning can successfully mitigate and those that it may inadvertently introduce. Through a systematic evaluation of 57 models, comparing 14 traditional models with 23 zero-shot LLMs and 20 fine-tuned variants, our results show that fine-tuning yields an average F1-score improvement of ＋15.1%. The best-performing model, a fine-tuned Mistral-7B, achieves an F1-score of 86.0%, although our analysis reveals that fine-tuning fails to produce meaningful performance gains in several scenarios.

Synthetic Data Generation (SDG) is expanding rapidly, yet existing surveys differ widely in scope and methodological quality. This tertiary study systematically searched four major scholarly databases (2015-2025) and, after PRISMA screening and DARE-4 appraisal, 1 identified 17 eligible secondary studies. The evidence reveals a strong concentration in healthcare (58.8% of surveys), limited coverage of non-health domains, and inconsistent reporting of evaluation protocols (e.g., incomplete specification of metrics, data splits, baselines, or evaluation scripts). Fidelity and downstream utility dominate assessment practices, whereas privacy and diversity remain under-examined. Only 4 of 17 surveys provide any reproducibility artefacts. By consolidating these findings, we propose a compact, domain-agnostic evaluation baseline and highlight structural gaps in transparency, domain breadth, and methodological consistency. The study offers actionable guidance for strengthening reproducibility and broadening the evidential foundations of SDG research.

In this paper, we present a stock market forecasting framework that integrates sentiment analysis of news headlines associated with individual stocks performed by Large Language Models (LLMs) with economic indicators. Specifically, we infer multi-class labels and continuous polarity scores by using LLMs (i.e., Llama, Vicuna, and Mistral) under zero-shot settings from news content. These sentiment signals are combined with historical price data and economic indicators and fed as input to different deep learning models (i.e., Long Short-Term Memory (LSTM) networks, Generative Adversarial Networks (GANs), and Transformers) for predicting stock closing prices. We evaluate the proposed framework on a dataset comprising historical financial time series and 8652 news articles related to 47 meme stocks listed on major stock exchanges, covering the period from January 2019 to December 2021. While Vicuna delivers the quickest sentiment processing (around faster than Llama), it achieves the greatest hallucination rate ( ). Among the forecasting models, Transformer architectures offer improved Mean Absolute Percentage Error (MAPE) scores ( – ) while LSTM requires minimal computational resources for training (up to ).

Designing and testing modern network systems is a complex and costly process, particularly during the testing phase, where time constraints and unpredictability often lead to inflated development efforts. In this paper, we propose a novel methodology that leverages natural language comments from developer commits to optimize test execution. By extracting meaningful insights from comments, we construct a prioritized task list, focusing on those tasks that are most likely to fail. This approach streamlines the testing workflow, accelerates execution, and reduces overall development costs. Our pipeline combines transformer-based models for semantic understanding with unsupervised anomaly detection methods, including clustering algorithms, autoencoders (AE), and variational autoencoders (VAE), to identify failure-prone scenarios without requiring labeled data. Applied to a real-world industrial dataset, the AE model achieves an F1score of 0.838, and the VAE follows closely with 0.805, significantly outperforming traditional clustering approaches. These results highlight the effectiveness of leveraging commit metadata for intelligent test prioritization and demonstrate the potential for scalable improvements in continuous integration pipelines.

In this paper, we present an analysis of the dual role of Large Language Models (LLMs) in the context of multimedia security, examining how these models can both strengthen defensive capabilities and introduce new vulnerabilities within modern content ecosystems. As LLMs increasingly interface with multimedia workflows-ranging from text-image generation pipelines to cross-modal retrieval and content moderation-their susceptibility to attacks such as data leakage, prompt injection, and jailbreaks raises critical concerns for the integrity and trustworthiness of multimedia platforms. We focus our analysis on privacy-oriented threats, including model inversion and membership inference, and discuss their implications for systems that process or generate multimedia content. Building on this risk landscape, we outline principles for improving the robustness and resilience of LLM-driven multimedia applications, highlighting strategies suited to in-the-wild threat scenarios. Indeed, we explore the constructive application of LLMs within cybersecurity frameworks, such as the Cyber Kill Chain, demonstrating how these models can be leveraged for threat detection, risk assessment, and automated defensive operations.

Online property booking platforms are widely used and rely heavily on consistent, up-to-date information about accommodation facilities, often sourced from third-party providers.However, these external data sources are frequently affected by incomplete or inconsistent details, which can frustrate users and result in a loss of market.In response to these challenges, we present an industrial case study involving the integration of Large Language Models (LLMs) into CALEIDOHOTELS, a property reservation platform developed by FERVENTO.We evaluate two well-know LLMs in this context: Mistral 7B, fine-tuned with QLoRA, and Mixtral 8x7B, utilized with a refined system prompt.Both models were assessed based on their ability to generate consistent and homogeneous descriptions while minimizing hallucinations.Mixtral 8x7B outperformed Mistral 7B in terms of completeness (99.6% vs. 93%), precision (98.8% vs. 96%), and hallucination rate (1.2% vs. 4%), producing shorter yet more concise content (249 vs. 277 words on average).However, this came at a significantly higher computational cost: 50GB VRAM and $1.61/hour versus 5GB and $0.16/hour for Mistral 7B.Our findings provide practical insights into the trade-offs between model quality and resource efficiency, offering guidance for deploying LLMs in production environments and demonstrating their effectiveness in enhancing the consistency and reliability of accommodation data.

Integrating physical and cyber realms, Cyber–Physical Systems (CPSs) expand the potential attack surface for intruders. Given their deployment in critical infrastructures like Industrial Control Systems (ICSs), ensuring robust security is imperative. Current research has developed various Intrusion Detection techniques to identify and counter malicious activities. However, traditional methods often encounter challenges in detecting several attack types due to reliance on a single data source such as time series data from sensors and actuators. In this study, we meticulously design advanced Deep Learning (DL) anomaly-based techniques trained on either sensor/actuator data or network traffic statistics in an unsupervised setting. We evaluate these techniques on network and physical data collected concurrently from a real-world CPS. Through meticulous hyperparameter tuning, we identify the optimal parameters for each model and compare their efficiency and effectiveness in detecting different types of attacks. In addition to demonstrating superior performance compared to various baselines, we showcase the best model for each data source. Eventually, we show how utilizing diverse data sources can enhance cyber-threat detection, recognizing different kinds of attacks. • We design DL anomaly-based techniques trained on either sensor or network data. • We evaluate techniques on different data sources collected from a real-world CPS. • After hyperparameter tuning, we compare the models’ attack detection ability. • We demonstrate superior performance compared to state-of-the-art base-lines. • We show how utilizing diverse data sources can enhance cyber-threat detection.

The protection of Cyber–Physical Systems (CPSs) from cybersecurity threats is essential to ensure the resilience and safety of critical infrastructures. Anomaly detection approaches for CPSs proposed in the literature use either network data or data from sensors/actuators as inputs, often failing to detect attacks that affect only specific components. In this paper, we propose a novel two-stage framework for threat detection in CPSs. This framework integrates anomaly detection models that operate on both network and physical data, by leveraging a decision fusion technique to combine the outputs into a coherent decision. To assess the effectiveness of the framework, we employ an unlabeled release of a real-world dataset, integrating network traffic with sensors/actuators data. Additionally, we offer explicit labeling rules to ensure reproducibility. The results demonstrate that our approach substantially improves CPSs security, efficiently identifying subtle attacks that can evade traditional methods relying on a single data source. In particular, we show that integrating both physical and network data improves the F1 score by approximately 10% compared to using just network data, and by nearly 30% compared to using just physical data. • We propose a framework for CPS threat detection using both network and physical data. • We integrate diverse deep learning models to analyze each data source independently. • We design a decision fusion technique to integrate network and sensor readings. • We achieve a 10% F1 improvement over network data and a 30% over sensor data. • Our method detects threats in 2–3 s, enabling quick response to CPS attacks.

Chronicles of the last few years show that industrial Cyber-Physical Systems are the target of dangerous cyber-attacks and face multiple threats. It is important to react as promptly as possible to such attacks and take proper countermeasures. Anomaly detection is a key activity in a Cyber-Physical System’s defense strategy. It involves analyzing sensor data, modeled as a Multivariate Time Series, to identify deviations from expected behavior, that may indicate potential cyber threats or attacks. In this paper, we design a novel framework integrating spatial and temporal modules to unveil spatio-temporal dependencies within sensor data in Cyber-Physical Systems to detect possible intrusions. We propose a novel strategy based on time series correlation to build a graph minimizing the number of sensors’ connections to unveil spatial dependencies between multimodal time series. The prediction and reconstruction losses are then leveraged to detect anomalies. The proposed framework has been evaluated on a real-world Cyber-Physical System, on which we evaluated both the efficacy and efficiency with respect to different competing approaches. The experimental analysis shows that the proposed framework outperforms eight state-of-the-art ones by increasing the precision of 0.59% while reducing both the training time (21.05%) for each epoch and memory occupation (77.8%) with respect to the best competitor in the literature. These characteristics make it particularly suitable for industrial environments that need periodic reconfigurations.

The high volatility and non-linear dynamics of the stock market make forecasting stock behavior increasingly challenging. This task is further influenced by a variety of contextual factors, including news content and social media data. However, several challenges remain unresolved, stemming from the influence of textual content, specifically from social networks, that are often characterized by dynamic interactions and non-standard language. In this paper, we design a two-step framework for the prediction of stock behavior by integrating contextual features inferred from social networks with historical financial data. We first build a chain of text composed of posts, and the related daily comments that are summarized and analyzed through Large Language Models to infer features in terms of sentiment scores and a buy/sell indicator. Successively, these social contextual features and historical financial data are fed as input to a forecasting module composed of different deep learning models (i.e., Long Short-Term Memory, Temporal Convolutional Networks, Transformers, and Generative Adversarial Networks). The proposed framework has been evaluated on a NASDAQ-listed stock dataset, encompassing more than 4 millions of social content and historical financial data over a period of six years of analysis. We show an improvement of up to 23.27% in Root Mean Square Error and 23.70% in Mean Average Error for forecasting stock behavior compared to various baselines. Finally, we compare these results with respect to those achieved using Large Language Models as predictors under zero-shot scenarios, achieving lower results (up to 68.1%) in addressing the forecasting task.

In the past years, the widespread diffusion of Artificial Intelligence (AI) in the finance domain transformed different services, with particular attention to the stock market. Although different AI-based approaches have been proposed for stock forecasting, they are focused on news content or sentiment without considering fundamental features and vice versa. In turn, other approaches rely on handmade rules or ones based on technical indicators for providing advice without considering contextual information that can strongly affect the stock market. In this paper, we propose an Advisor Neural Network framework using Long Short-Term Memory (LSTM)-based Informative Stock Analysis for Daily investment Advice. Specifically, the forecasting unit relies on a LSTM-based model, which combines technical indicators, contextual information, and financial data for stock forecasting. Successively, the advice unit provides next-day advice based on predicted information in conjunction with the proposed Heuristic Stocks Selection algorithm. This framework has been evaluated on the Stock and Cryptocurrencies markets, considering a subset of 417 stocks and 67 cryptocurrencies over three years, respectively. We compared the proposed framework with several state-of-the-art approaches, showing how it outperforms the baseline in both markets. Furthermore, we achieved a financial gain greater than 41%, despite the downward trend of the NASDAQ market in the quarter under review, and we obtained a 39.38% return on investment for the Cryptocurrencies market.

Context: In recent years, the code translation task has arisen as one of the major software issues in maintaining software quality during migration over complex infrastructure. This task involves human subjects with different background knowledge and could introduce errors due to the semantic gap between the programming languages and the complexity of the task. Generative Artificial Intelligence (AI) showed good capabilities in code generation, albeit this is highly dependent on the human factor . Objective: This paper investigates, from the human perspective, the use of three Generative AI tools (ChatGPT, Google Bard, and GitHub Copilot) in the context of translation tasks from code written in query languages to code written in framework-specific code languages, specifically focused on SQL dialects and PySpark. This translation is especially crucial during the migration from centralized architectures to cloud-based architectures. Methods: We evaluate the usefulness of these tools, the quality of the generated code, and their impact on performance. The models are tested with queries of various type in three different SQL dialects considering three usage scenarios of increasing complexity. It involves 15 participants with diverse programming backgrounds, who aim to solve tasks by interacting multiple times with the tools and manually changing the code. Results: The findings show a positive performance, demonstrating their reliability in generating coherent translations, achieving 100% precision in most tasks with a slight decrease in more complex scenarios, and producing well-documented code, with a response time of under 2 min, with Google Bard responding 50% faster than the others. Conclusion: In conclusion, this paper establishes a methodology and both quantitative and qualitative metrics for evaluating how generative AI tools streamline code translation, shifting the emphasis from production to refinement. It underscores the importance of continuously improving these tools to integrate them into developers’ workflows and to provide guidelines for intelligent use.

The spread of misinformation and fake news on digital platforms poses significant societal challenges, underscoring the need for robust detection. Multi-task learning leverages relationships among disinformation-related tasks (e.g., stance detection, rumor classification) to enhance detection; however, it risks negative transfer , potentially degrading performance instead of achieving positive transfer . In this paper, we systematically investigate the mechanisms underlying positive and negative transfers across a comprehensive set of disinformation-related tasks, including Sentiment Analysis (SA), Fake News Detection (FND), Stance Detection (SD), and Topic Detection (TD). Specifically, we pioneer the use of explanations to uncover the differences between models trained under single-task and multi-task settings. Our results reveal instances of positive transfer across several task combinations, with multi-task learning yielding performance improvements of 3.26%, 6.57%, and 0.62% for SA, FND and TD tasks, respectively. Furthermore, when comparing explanations of single-task and multi-task models, we find that positive transfer refines the knowledge that can already be learnt in single-task settings by incorporating additional patterns from other tasks. Conversely, negative transfer significantly undermines models' knowledge to the extent that their explanations are equivalent to a random perturbation of the explanations generated by their single-task counterparts.

In last years, social lending platforms have been increasingly used as virtual environments where borrowers can directly interact with lenders without any intermediary. As a result, a reliable credit scoring strategy, i.e., assessing whether a client is able to fully repay a loan, became of utmost importance to reduce the risk of not repaying the lenders. In this context, machine learning tools are being increasingly adopted to design automatic credit scoring systems but the data imbalance problem still penalizes their predictive performance, i.e., the greatest majority of clients can afford the repayment and learning to classify ”bad” borrowers depends on few instances where the loan was not paid back. In this paper, we target the data imbalance problem and propose a novel data augmentation strategy to improve the predictive performance of credit scoring models. The proposed methodology performs data augmentation by injecting synthetic instances in the dataset generated along the decision boundary of the decision model. We assessed the effectiveness of the proposed augmentation strategy on a million-scale dataset from Lending Club, the largest Social Lending platform, and found that it improves the performance of several classification models, also in comparison to other state-of-the-art approaches.

Nowadays, our security and privacy are strongly threatened by malware programs which aim to steal our confidential data and make our systems out of service, among other things. While traditional signature-based malware detection methods or statistical analysis have proven to be ineffective and time-consuming, recently data-driven Artificial Intelligence (AI) techniques, i.e. Machine Learning (ML) and Deep Learning (DL) approaches, have been successfully applied leveraging the behaviour of malware in terms of API calls, and achieving promising performances. However, their black-box behavior leads to a lack of explainability thus preventing their application in real world scenarios. In light of this, eXplainable Artificial Intelligence (XAI) methodologies and tools can be effectively embedded within an AI-based malware detection process in order to make more understandable the produced results. In this paper, we propose a XAI framework for behavioral malware detection problems and evaluate the usefulness of four XAI methods (SHAP, LIME, LRP and Attention mechanism) on three datasets with different size, sequence length and number of classes, by which we could evaluate the strengths and weaknesses – from effectiveness and efficiency point of views – of recurrent deep architectures (i.e. Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) models), and their applicability in the modern Cyber Security (CS) scenarios.

Conducting a maturity assessment allows companies to measure their readiness in implementing novel technologies. However, this task is challenging due to the multidimensional, complex, unpredictable, and non-linear nature of innovation. In this paper, we introduce an innovative approach to maturity assessment that enables both intra- and inter-company analysis. Our approach evaluates a company’s absolute maturity score concerning a specific technology or area. By leveraging a Natural Language Processing pipeline applied to a semi-structured questionnaire we extract popular concepts from the answers and present them to a human expert for analysis. The expert can refine the analysis by adding or removing concepts as needed. Subsequently, we compute a similarity metric for each answer to determine a company’s maturity in specific concepts. The output of our analysis is presented through human-readable plots, offering clear insights into the internal maturity level of the company and allowing for a comparison with competitors across the chosen concepts. To demonstrate the capabilities of our method, we provide a running example showcasing both quantitative and qualitative results of the analysis. Our approach demonstrates efficiency, with preprocessing completed in 1.967±0.758 s, and information extraction in 0.074±0.017 s on average, excluding human intervention time, and requiring low hardware resources.

In this paper, we address issues associated with the effective management of handover events in satellite-enabled 5G network infrastructures. Namely, we devise a strategy for dynamically allocating 5G gNB available resources in the presence of a constellation of LEO satellites, based on several parameters collected and dispatched by an ad hoc orchestration platform. We propose to leverage a Combinatorial Multi-Armed Bandit approach to design a resource allocation game that helps make decisions over time under uncertainty conditions related to the incidence of several factors that can determine the quality of experience perceived on the user equipment side. The introduced approach does represent a pioneering one, since it allows us to model a joint optimization task as a competitive game in which agents typically share resources with other agents instead of occupying them exclusively. The designed task allows to dynamically enable and disable channels, taking care of the relationships with the lower layer, transparently managing the required handover operations, and considering possible interference with other channels. We discuss an implementation of the proposed solution in a simulated environment. We also analyze the performance it attains, by measuring both its efficiency and efficacy in a trial setup reproducing a scenario with near real-time temporal requirements. Results show that the proposed approach has a linear trend in terms of running time with respect to the number of user equipments and gNbs involved, while achieving a sub-optimal solution of the handover task in around 20–30 rounds.

Ubiquitous utilization of Information and Communication Technologies in modern manufacturing plants has transformed them into Cyber-Physical Systems (CPSs), making them susceptible to cyber-attacks, which can have huge economic and social impact. In this paper, we focus on the security issues that may affect the Industrial Control System (ICS) supervising an industrial establishment. The purpose of this work is to provide readers with an up-to-date view of the methodologies that current literature suggests as the most appropriate to defend an ICS against cyber-attacks. We firstly provide a classification of existing attacks according to the methodology used by the attacker. Subsequently, we propose a classification of defensive countermeasures in Model and Artificial Intelligence-based approaches by analyzing most recent research contributions. Furthermore, we describe the most used datasets in literature that are available to researchers and practitioners. We conclude the paper by discussing today's open issues, which need to be addressed to cope with the increasing complexity of ICS security.

Although in the last decade several fact-checking organizations have emerged to verify misinformation, fake news has continued to proliferate, especially through social media platforms. Even though adopting improved detection strategies is of utmost importance, the fact-checking process could be optimised by verifying whether a claim has been previously fact-checked. Despite some ad-hoc information retrieval approaches having been recently proposed, the utility of modern (neural) retrieval systems have not been investigated yet. In this paper, we consider the standard two-phases retriever-reranker architecture and benchmark different state-of-the-art techniques from the information retrieval and Q&A literature. We design several experiments on a real-world Twitter dataset to analyse the efficiency and the effectiveness of the benchmark approaches. Our results show that combining standard and neural approaches is the most promising research direction to improve retrievers performance and that complex (neural) rerankers might still be efficient in practice since there is no need to process a high number of documents to improve ranking performance.

Recent years have seen an exponential growth (+98% in 2022 w.r.t. the previous year) of the number of research articles in the few-shot learning field, which aims at training machine learning models with extremely limited available data. The research interest toward few-shot learning systems for Named Entity Recognition (NER) is thus at the same time increasing. NER consists in identifying mentions of pre-defined entities from unstructured text, and serves as a fundamental step in many downstream tasks, such as the construction of Knowledge Graphs, or Question Answering. The need for a NER system able to be trained with few-annotated examples comes in all its urgency in domains where the annotation process requires time, knowledge and expertise (e.g., healthcare, finance, legal), and in low-resource languages. In this survey, starting from a clear definition and description of the few-shot NER (FS-NER) problem, we take stock of the current state-of-the-art and propose a taxonomy which divides algorithms in two macro-categories according to the underlying mechanisms: model-centric and data-centric. For each category, we line-up works as a story to show how the field is moving toward new research directions. Eventually, techniques, limitations, and key aspects are deeply analyzed to facilitate future studies.

In Biomedical Named Entity Recognition (BioNER), the use of current cutting-edge deep learning-based methods, such as deep bidirectional transformers (e.g. BERT, GPT-3), can be substantially hampered by the absence of publicly accessible annotated datasets. When the BioNER system is required to annotate multiple entity types, various challenges arise because the majority of current publicly available datasets contain annotations for just one entity type: for example, mentions of disease entities may not be annotated in a dataset specialized in the recognition ofdrugs, resulting in a poor ground truth when using the two datasets to train a single multi-task model. In this work, we proposeTaughtNet, a knowledge distillationbased framework allowing us to fine-tune a single multi-task student model by leveraging both the ground truth and the knowledge of single-taskteachers. Our experiments on the recognition of mentions of diseases, chemical compounds and genes show the appropriateness and relevance of our approach w.r.t. strong state-of-the-art baselines in terms of precision, recall and F1 scores. Moreover, TaughtNet allows us to train smaller and lighter student models, which may be easier to be used in real-world scenarios, where they have to be deployed on limitedmemory hardware devices and guarantee fast inferences, and shows a high potential to provide explainability. We publicly release both our code on github1 and our multi-task model on the huggingface repository.

In recent years, we have witnessed an ever wider spread of multimedia streaming platforms (e.g., Netflix, Spotify, and Amazon). Hence, it has become more and more essential to provide such systems with advanced recommendation facilities, in order to support users in browsing these massive collections of multimedia data according to their preferences and needs. In this context, the modeling of entities and their complex relationships (e.g., users listening to topic-based songs or authors creating different releases of their lyrics) represents the key challenge to improve the recommendation and maximize the users' satisfaction. To this end, this is the first study to leverage the high representative power of hypergraph data structures in combination with modern graph machine learning techniques in the context of music recommendation. Specifically, we propose hypergraph embeddings for music recommendation (HEMR), a novel framework for song recommendation based on hypergraph embedding. The hypergraph data model allows us to represent seamlessly all the possible and complex interactions between users and songs with the related characteristics; meanwhile, embedding techniques provide a powerful way to infer the user-song similarities by vector mapping. We have experimented the effectiveness and efficiency of our approach with respect to the state-of-the-art most recent music recommender systems, exploiting the Million Song dataset. The results show that HEMR significantly outperforms other state-of-the-art techniques, especially in scenarios where the cold-start problem arises, thus making our system a suitable solution to embed within a music streaming platform.

Federated learning has been increasingly studied to cope with the scalability and privacy issues characterizing current and upcoming large-scale infrastructures, such as the Internet of Things, 5G networks, vehicular applications, and so on. This approach partitions the data storage and AI model training in a series of local learners, whose results are aggregated by a central server and then redistributed back to the learners to have a trained global model. However, despite avoiding outsourcing sensitive data to cloud-hosted services and fragmenting the workload for data processing, such a decentralized learning approach lays the overall solution open to various kinds of attacks, able to fully compromise the accuracy of the obtained global model. This study aims to quantitatively assess the impact of two widely-recognized attacks against federated learning and propose a tentative protection means by using blockchain.

Nowadays, really huge volumes of fake news are continuously posted by malicious users with fraudulent goals thus leading to very negative social effects on individuals and society and causing continuous threats to democracy, justice, and public trust. This is particularly relevant in social media platforms (e.g., Facebook, Twitter, Snapchat), due to their intrinsic uncontrolled publishing mechanisms. This problem has significantly driven the effort of both academia and industries for developing more accurate fake news detection strategies: early detection of fake news is crucial. Unfortunately, the availability of information about news propagation is limited. In this paper, we provided a benchmark framework in order to analyze and discuss the most widely used and promising machine/deep learning techniques for fake news detection, also exploiting different features combinations w.r.t. the ones proposed in the literature. Experiments conducted on well-known and widely used real-world datasets show advantages and drawbacks in terms of accuracy and efficiency for the considered approaches, even in the case of limited content information.

Hard disk drive failures are one of the most common causes of service downtime in data centers. Predictive maintenance techniques have been adopted to extend the Remaining Useful Life (RUL) of these drives, and minimize service shortage and data loss. Several approaches based on machine and deep learning techniques have been proposed to address these issues, mostly exploiting models based on Self-Monitoring analysis and Reporting Technology (SMART) attributes. While these models have proven to be reliable, their performance is affected by the lack of information about the proximity of disk failure in time. Moreover, many of these techniques are sensitive to the highly unbalanced nature of existing data-sets, in terms of good to failed hard disk ratio. In this article, we propose a LSTM (Long Short Term Memory)-based model combining SMART attributes and temporal analysis for estimating a hard drive health status according to its time to failure. Our approach outperforms state-of-the-art methods when evaluated on two data-sets, one containing hourly samples from 23395 disks and the other reporting daily samples from 29878‬ disks. Experimental results showed that our approach is well suited to data-sets with different sampling periods, being able to predict hard drive health status up to 45 days before failure.

Past research has shown that over 8 months may elapse between the time when a network is compromised and the time the attack is discovered. During this long gap, attackers can steal valuable intellectual property from the victim. The recent FORGE system [8] has suggested that automatically generating fake—but believable—versions of documents can delay the attacker, cost him money, and increase his uncertainty. However, in order to generate fakes, FORGE only modifies the textual component of the document in question. But in the real world, documents consist of many non-textual components such as charts, equations, formulas, diagrams, and tables. We propose the concept of a Probabilistic Logic Graph (PLG) and show that PLGs provide a single, unified framework within which the different parts of a document can be expressed. We then define the problem of generating, for a given PLG representation of a document, a set of fake yet highly believable PLGs (i.e., documents), so that an attacker looking at them (both the original and the fake ones) cannot easily identify the original document. We show that the problem of generating fake PLGs is intractable—but we propose an approximation algorithm that solves it efficiently. We evaluate the use of PLGs over a corpus of patents and show that our fakes can effectively deceive an adversary.

Social Networks represent a cornerstone of our daily life, where the so-called social reviewing systems (SRSs) play a key role in our daily lives and are used to access data typically in the form of reviews. Due to their importance, social networks must be trustworthy and secure, so that their shared information can be used by the people without any concerns, and must be protected against possible attacks and misuses. One of the most critical attacks against the reputation system is represented by mendacious reviews. As this kind of attacks can be conducted by legitimate users of the network, a particularly powerful solution is to exploit trust management, by assigning a trust degree to users, so that people can weigh the gathered data based on such trust degrees. Trust management within the context of SRSs is particularly challenging, as determining incorrect behaviors is subjective and hard to be fully automatized. Several attempts in the current literature have been proposed; however, such an issue is still far from been completely resolved. In this study, we propose a solution against mendacious reviews that combines fuzzy logic and the theory of evidence by modeling trust management as a multicriteria multiexpert decision making and exploiting the novel concept of time-dependent and content-dependent crown consensus. We empirically proved that our approach outperforms the main related works approaches, also in dealing with sockpuppet attacks.

Satellite communications (SatCom) have a role of advanced service enablers in the new virtual networks, following 3GPP specifications. Specifically, the satellite peculiar characteristics are of paramount importance to dynamically activate capabilities, such as multicast and broadcast channels, sudden traffic offloading, capacity bonding, and cost-efficient coverage of uncovered areas. A key aspect to achieve a seamless and efficient integration between satellite and terrestrial infrastructures is to make satellite resource management dynamic and with a centralised control of a single orchestrator that has visibility of the end-to-end network. Considering the adoption of Software Defined Networking (SDN) and Network Function Virtualization (NFV) paradigms and the upcoming 5thgeneration of mobile communications (5G), this paper presents the ongoing work within the European Space Agency (ESA) ANChOR project, whose main output will be a data-driven Network Controller and Orchestrator for SatCom networks. This tool will be based on Artificial Intelligence (AI) / Machine Learning (ML)-based techniques to properly allocate resources exploiting some feedback knowledge of the network and it aims to support different services over 5G integrated satellite-terrestrial networks.

The pervasive diffusion of Social Networks (SN) produced an unprecedented amount of heterogeneous data. Thus, traditional approaches quickly became unpractical for real life applications due their intrinsic properties: large amount of user-generated data (text, video, image and audio), data heterogeneity and high speed generation rate. More in detail, the analysis of user generated data by popular social networks (i.e Facebook (https://www.facebook.com/), Twitter (https://www.twitter.com/), Instagram (https://www.instagram.com/), LinkedIn (https://www.linkedin.com/)) poses quite intriguing challenges for both research and industry communities in the task of analyzing user behavior, user interactions, link evolution, opinion spreading and several other important aspects. This survey will focus on the analyses performed in last two decades on these kind of data w.r.t. the dimensions defined for Big Data paradigm (the so called Big Data 6 V's).

With the recent COVID-19 outbreak, we have assisted to the development of new epidemic models or the application of existing methodologies to predict the virus spread and to analyze how the different lock-down strategies can effectively influence the epidemic diffusion. In this paper, we propose a novel machine learning based framework able to estimate the parameters of any epidemiological model, such as contact rates and recovery rates, based on static and dynamic features of places. In particular, we model mobility data through a graph series whose spatial and temporal features are investigated by combining Graph Convolutional Neural Networks (GCNs) and Long short-term memories (LSTMs) in order to infer the parameters of SIR and SIRD models. We evaluate the proposed approach using data related to the COVID-19 dynamics in Italy and we compare the forecasts of the trained model with available data about the epidemic spread.

Today, major corporations and government organizations must face the reality that they will be hacked by malicious actors. In this paper, we consider the case of defending enterprises that have been successfully hacked by imposing additional a posteriori costs on the attacker. Our idea is simple: for every real document $d$ d , we develop methods to automatically generate a set $Fake(d)$ F a k e ( d ) of fake documents that are very similar to $d$ d . The attacker who steals documents must wade through a large number of documents in detail in order to separate the real one from the fakes. Our $\mathsf {FORGE}$ FORGE system focuses on technical documents (e.g., engineering/design documents) and involves three major innovations. First, we represent the semantic content of documents via multi-layer graphs (MLGs). Second, we propose a novel concept of “meta-centrality” for multi-layer graphs. A meta-centrality (MC) measure takes a classical centrality measure (for ordinary graphs, not MLGs) as input, and generalizes it to MLGs. The idea is to generate fake documents by replacing concepts on the basis of meta-centrality with related concepts according to an ontology. Our third innovation is to show that the problem of generating the set $Fake(d)$ F a k e ( d ) of fakes can be viewed as an optimization problem. We prove that this problem is NP-complete and then develop efficient heuristics to solve it in practice. We ran detailed experiments on two datasets: one a panel of 20 human subjects, another with a panel of 10. Our results show that $\mathsf {FORGE}$ FORGE generates highly believable fakes.

The uncontrolled growth of fake news creation and dissemination we observed in recent years causes continuous threats to democracy, justice, and public trust. This problem has significantly driven the effort of both academia and industries for developing more accurate fake news detection strategies. Early detection of fake news is crucial, however the availability of information about news propagation is limited. Moreover, it has been shown that people tend to believe more fake news due to their features [10]. In this paper, we present our framework for fake news detection and we discuss in detail an approach based on deep learning that we implemented by using Google Bert features. Our experiments conducted on two well-known and widely used real-world datasets suggest that our method can outperform the state-of-the-art approaches and allows fake news accurate detection, even in the case of limited content information.

We demonstrate GraphDBLP, a tool to allow researchers for querying the DBLP bibliography as a graph. The DBLP source data were enriched with semantic similarity relationships computed using wordembeddings. A user can interact with the system either via a Web-based GUI or using a shell-interface, both provided with three parametric and pre-defined queries. GraphDBLP would represent a first graph-database instance of the computer scientist network, that can be improved through new relationships and properties on nodes at any time, and this is the main purpose of the tool, that is freely available on Github. To date, GraphDBLP contains 5+ million nodes and 24+ million relationship.

Recommender systems are increasingly playing an important role in our life, enabling users to find “what they need” within large data collections and supporting a variety of applications, from e-commerce to e-tourism. In this paper, we present a Big Data architecture supporting typical cultural heritage applications. On the top of querying, browsing, and analyzing cultural contents coming from distributed and heterogeneous repositories, we propose a novel user-centered recommendation strategy for cultural items suggestion. Despite centralizing the processing operations within the cloud, the vision of edge intelligence has been exploited by having a mobile app (Smart Search Museum) to perform semantic searches and machine-learning-based inference so as to be capable of suggesting museums, together with other items of interest, to users when they are visiting a city, exploiting jointly recommendation techniques and edge artificial intelligence facilities. Experimental results on accuracy and user satisfaction show the goodness of the proposed application.

Summary Nowadays, social networks provide users an interactive platform to create and share heterogeneous content for a lot of different purposes (eg, to comment events and facts, and express and share personal opinions on specific topics), allowing millions of individuals to create online profiles and share personal information with vast networks known and sometimes also unknown people. Knowledge about users, content, and relationships in a social network may be used for an adversary attack of some victims easily. Although a number of works have been done for data privacy preservation on relational data, they cannot be applied in social networks and in general for big data analytics. In this paper, we first propose a novel data model that integrates and combines information on users belonging to 1 or more heterogeneous online social networks, together with the content that is generated, shared, and used within the related environments, using an hypergraph data structure; then we implemented the most diffused centrality measures and also introduced a new centrality measure—based on the concept of “neighborhood” among users—that may be efficiently applied for a number of data privacy issues, such as lurkers and neighborhood attack prevention, especially in “interest‐based” social networks. Some experiments using the Yelp dataset are discussed.

Despite the great amount of research done in the Online Social Networks (OSNs) field, only few works have investigated the use of multimedia data in such realm. Instead, it is the authors' opinion that a novel data model that takes into account the intrinsic characteristics of multimedia may be of great help in managing Multimedia OSNs for providing more effective algorithms. In this paper, we describe a novel OSN data model that supports easy management of multimedia content in a unique framework, providing a more effective and efficient mechanism for data and information management in a variety of applications, especially for Influence Analysis aims.

Nowadays, the need for well-structured ontologies in the medical domain is rising, especially due to the significant support these ontologies bring to a number of groundbreaking applications, such as intelligent medical diagnosis system and decision-support systems. Indeed, the considerable production of clinical data belonging to restricted sub domains has stressed the need for efficient methodologies to automatically process enormous amounts of un-structured, domain specific information in order to make use of the knowledge these data provide. In this work, we propose a lexicon-grammar based methodology for efficient information extraction and retrieval on unstructured medical records in order to enrich a simple ontology descriptive of such a kind of documents. We describe the NLP methodology for extracting RDF triples from unstructured medical records, and show how an existing ontology built by a domain expert can be populated with the set of triples and then enriched through its linking to external resources.

In this paper, we present a framework supporting information retrieval over corpora of documents using an automatic sematic query expansion approach. The main idea is to expand the set of words used as query terms exploiting the notion of semantic similarity between the concepts related to the search terms. We leverage existing lexical resources and similarity metrics computed among terms to generate - by a proper mapping into a vectorial space - an index for the fast retrieval of a set of terms "semantically correlated" to a given query term. The vector of expanded terms is then exploited in the query stage to retrieve documents that are significantly related to specific combinations of the query terms. Preliminary experimental results concerning efficiency and effectiveness of the proposed approach are reported and discussed.