The Network Digital Twin (NDT) developed in the B5GEMINI project is presented in this article, highlighting its architecture, objectives, functionalities, and practical applications. The design of the NDT architecture is detailed, including the establishment of the foundational infrastructure, developed as part of B5GEMINI-INFRA. The integration of artificial intelligence techniques for network management tasks within B5GEMINI-AIUC is illustrated with relevant use cases, such as the detection of cybersecurity attacks and the simulation and optimization of virtual reality applications, to demonstrate the usefulness and potential of the proposed NDT solution. The platform enables controlled experimentation and data collection for training Machine Learning (ML) models, addressing challenges associated with realistic network traffic datasets and cybersecurity experiments without disrupting live networks. The infrastructure supporting the NDT allows for creating virtual scenarios, isolating traffic between experiments, on-demand traffic generation, and capture, ensuring repeatability and enabling evaluation of different detection and mitigation tools under identical conditions. Additionally, an in-depth use case focusing on ML-based detection of a simulated denial of service attack through DNS over HTTPS within a 5G network framework showcases the NDT’s potential to provide a secure environment for testing and validating ML-based solutions without disrupting live networks.

The increasing occurrence of Harmful Algal Blooms (HABs) in water systems poses significant challenges to ecological health, public safety, and economic stability globally. Deep Learning (DL) models, notably Convolutional Neural Networks (CNN) and Long-Short Term Memory (LSTM), have been widely employed for HAB prediction. However, the emergence of state-of-the-art multi-horizon forecasting DL architectures such as Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS) provides a novel solution for long-term HAB prediction. This study compares the performance of N-BEATS with LSTM and CNN models using high temporal granularity water quality data from As Conchas reservoir (NW Spain) to forecast chlorophyll-a (Chl-a) concentrations, a key indicator of HABs. The evaluation encompasses one-day and one-week prediction horizons, aligning with World Health Organization (WHO) HAB alert criteria. Results indicate that N-BEATS outperforms LSTM and CNN models for one-week predictions and when forecasting multiple consecutive days within a week. Furthermore, augmenting input data with additional variables does not significantly enhance predictive accuracy, challenging the assumption that complexity always improves model performance. The study also explores the transferability of trained models across different monitoring buoys within the same water body, emphasizing the adaptability and broad applicability of predictive models in diverse aquatic environments. This research underscores the potential of N-BEATS as a valuable tool for HAB prediction, particularly for longer-term forecasting.

Abstract The growing prevalence of deep neural networks (DNNs) across various fields raises concerns about their increasing energy consumption, especially in large data center applications. Identifying the best combination of optimization techniques to achieve maximum energy efficiency while maintaining system performance is challenging due to the vast number of techniques available, their complex interplay, and the rigorous evaluation required to assess their impact on the model. To address this gap, we propose an open-source methodological framework for the systematic study of the influence of various optimization techniques on diverse tasks and datasets. The goal is to automate experimentation, addressing common pitfalls and inefficiencies of trial and error, saving time, and allowing fair and reliable comparisons. The methodology includes model training, automatic application of optimizations, export of the model to a production-ready format, and pre- and post-optimization energy consumption and performance evaluation at inference time using various batch sizes. As a novelty, the framework provides pre-configured "optimization strategies" for combining state-of-the-art optimization techniques that can be systematically evaluated to determine the most effective strategy based on real-time energy consumption and performance feedback throughout the model life cycle. As an additional novelty, "optimization profiles" allow the selection of the optimal strategy for a specific application, considering user preferences regarding the trade-off between energy efficiency and performance. Validated through an empirical study on a DNN-based cyber threat detector, the framework demonstrates up to 82% reduction in energy consumption during inference with minimal accuracy loss.

Abstract A novel AGV (Automated Guided Vehicle) control architecture has recently been proposed where the AGV is controlled remotely by a virtual Programmable Logic Controller (PLC), which is deployed on a Multi-access Edge Computing (MEC) platform and connected to the AGV via a radio link in a 5G network. In this scenario, we leverage advanced deep learning techniques based on ensembles of N-BEATS (state-of-the-art in time-series forecasting) to build predictive models that can anticipate the deviation of the AGV’s trajectory even when network perturbations appear. Therefore, corrective maneuvers, such as stopping the AGV, can be performed in advance to avoid potentially harmful situations. The main contribution of this work is an innovative application of the N-BEATS architecture for AGV deviation prediction using sequence-to-sequence modeling. This novel approach allows for a flexible adaptation of the forecast horizon to the AGV operator’s current needs, without the need for model retraining or sacrificing performance. As a second contribution, we extend the N-BEATS architecture to incorporate relevant information from exogenous variables alongside endogenous variables. This joint consideration enables more accurate predictions and enhances the model’s overall performance. The proposed solution was thoroughly evaluated through realistic scenarios in a real factory environment with 5G connectivity and compared against main representatives of deep learning architectures (LSTM), machine learning techniques (Random Forest), and statistical methods (ARIMA) for time-series forecasting. We demonstrate that the deviation of AGVs can be effectively detected by using ensembles of our extended N-BEATS architecture that clearly outperform the other methods. Finally, a careful analysis of a real-time deployment of our solution was conducted, including retraining scenarios that could be triggered by the appearance of data drift problems.

Abstract With the latest advances in deep learning-based generative models, it has not taken long to take advantage of their remarkable performance in the area of time series. Deep neural networks used to work with time series heavily depend on the size and consistency of the datasets used in training. These features are not usually abundant in the real world, where they are usually limited and often have constraints that must be guaranteed. Therefore, an effective way to increase the amount of data is by using data augmentation techniques, either by adding noise or permutations and by generating new synthetic data. This work systematically reviews the current state of the art in the area to provide an overview of all available algorithms and proposes a taxonomy of the most relevant research. The efficiency of the different variants will be evaluated as a central part of the process, as well as the different metrics to evaluate the performance and the main problems concerning each model will be analysed. The ultimate aim of this study is to provide a summary of the evolution and performance of areas that produce better results to guide future researchers in this field.

Digital Twin Network (DTN) is a new technology that builds on the concept of Digital Twins (DT) to create a virtual representation of the physical objects of a telecommunications network. DTN bridges physical and virtual spaces to enable coordination and synchronization of physical parts while eliminating the need to directly interact with them. In this work, we present B5GEMINI a DTN for 5G and beyond networks that makes an extensive use of artificial intelligence (AI). First, we present the infrastructural and architectural components that support B5GEMINI. Next, we explore five paradigmatic use cases where AI can leverage B5GEMINI for building new AI-powered applications. Finally, we identify the main components of the AI ecosystem of B5GEMINI.

Service providers, 5G network operators and, more generally, vertical industries face today a dangerous shortage of highly skilled cybersecurity experts. Along with the escalation and growing sophistication of cyber-attacks, 5G networks require the training of skilled and highly competent cyber forces. To meet these requirements, the SPIDER cyber range focuses specifically on 5G, and is based on three pillars, (i) cyber security assessment, (ii) training cyber security teams to defend against complex cyber-attack scenarios, and (iii) evaluation of cyber risk. The SPIDER cyber range replicates a customized 5G network, enabling the execution of cyber-exercises that take advantage of hands-on interaction in real time, the sharing of information between participants, and the gathering of feedback from network equipment, as well as the development and adaptation of advanced operational procedures. This aims to help 5G security professionals improve their ability to collaboratively manage and predict security incidents, complex attacks, and propagated vulnerabilities. The SPIDER cyber range is validated in two relevant use case scenarios aimed at demonstrating, in a realistic, measurable, and replicable way the transformations SPIDER will bring to the cybersecurity industry.

Solving the convergence issues of Generative Adversarial Networks (GANs) is one of the most outstanding problems in generative models. In this work, we propose a novel activation function to be used as output of the generator agent. This activation function is based on the Smirnov probabilistic transformation and it is specifically designed to improve the quality of the generated data. In sharp contrast to previous works, our activation function provides a more general approach that deals not only with the replication of categorical variables but with any type of data distribution (continuous or discrete). Moreover, our activation function is derivable and therefore, it can be seamlessly integrated in the backpropagation computations during the GAN training processes. To validate this approach, we firstly evaluate our proposal on two different data sets: a) an artificially rendered data set containing a mixture of discrete and continuous variables, and b) a real data set of flow-based network traffic data containing both normal connections and cryptomining attacks. In addition, three publicly available data sets were added to the evaluation to generalize the obtained results. To evaluate the fidelity of the generated data, we analyze their results both in terms of quality measures of statistical nature and regarding the use of these synthetic data to feed a nested machine learning-based classifier. The experimental results evince a clear outperformance of a Wasserstein GAN network (WGAN) tuned with this new activation function with respect to both a naïve mean-based generator and a standard WGAN. The quality of the generated data allows to fully substitute real data with synthetic data for training the nested classifier without a significant fall in the obtained accuracy.

Machine Learning (ML) technologies applied to Cybersecurity, especially in the area of network cyber threat detection, are a promising choice, but they require additional research in the applicability of a wide range of available algorithms. Such algorithms usually require training using good-quality and quantitatively significant datasets, which are rarely publicly available. To this end, in this paper we describe a novel experimental framework, that we call the Mouseworld, that combines NFV and SDN to create an environment able to (1) blend and transmit real and synthetic traffic and (2) collect and label this traffic in order to be utilised for training and validating ML algorithms that will be applied to the detection of cybersecurity threats. The Mouseworld framework includes a set of traffic generation, collection and labelling modules, jointly with analytics and algorithm training and visualization components. The OSM open-source network orchestrator is utilized to control and manage the framework and to deploy the training and validation scenarios. We present a preliminary result on the area of Security threat detection as a demonstration of the framework viability.

Dimensionality reduction is often a crucial step for the successful application of machine learning and data mining methods. One way to achieve said reduction is feature selection. Due to the impossibility of labelling many data sets, unsupervised approaches are frequently the only option. The column subset selection problem translates naturally to this purpose, and has received consider able attention over the last few years, as it provides simple linear models for data reconstruction. Existing methods, however, often achieve approximation errors that are far from the optimum. In this paper we present a novel algorithm for column subset selection that consistently outperforms state-of-the-art methods in approximation error. We present a series of key derivations that allow an efficient implementation, making it comparable in speed and in some cases faster than other algorithms. We also prove results that make it possible to deal with huge matrices, which has strong implications for other algorithms of this type in the big data field. We validate our claimsthrough experiments on a wide variety of well-known data sets.

5G represents the next generation of communication networks and services, and will bring a new set of use cases and scenarios. These in turn will address a new set of challenges from the network and service management perspective, such as network traffic and resource management, big data management and energy efficiency. Consequently, novel techniques and strategies are required to address these challenges in a smarter way. In this paper, we present the limitations of the current network and service management and describe in detail the challenges that 5G is expected to face from a management perspective. The main contribution of this paper is presenting a set of use cases and scenarios of 5G in which machine learning can aid in addressing their management challenges. It is expected that machine learning can provide a higher and more intelligent level of monitoring and management of networks and applications, improve operational efficiencies and facilitate the requirements of the future 5G network.

The growth of the Internet has increased the need for scalable congestion control mechanisms in high speed networks. In this context, we propose a rate-based explicit congestion control mechanism with which the sources are provided with the rate at which they can transmit. These rates are computed with a distributed max-min fair algorithm, SLBN. The novelty of SLBN is that it combines two interesting features not simultaneously present in existing proposals: scalability and fast convergence to the max-min fair rates, even under high session churn. SLBN is scalable because routers only maintain a constant amount of state information (only three integer variables per link) and only incur a constant amount of computation per protocol packet, independently of the number of sessions that cross the router. Additionally, SLBN does not require processing any data packet, and it converges independently of sessions' RTT. Finally, by design, the protocol is conservative when assigning rates, even in the presence of high churn, which helps preventing link overshoots in transient periods. We claim that, with all these features, our mechanism is a good candidate to be used in real deployments.

The number of online real-time streaming services deployed over network topologies like P2P or centralized ones has remarkably increased in the recent years. This has revealed the lack of networks that are well prepared to respond to this kind of traffic. A hybrid distribution network can be an efficient solution for real-time streaming services. This paper contains the experimental results of streaming distribution in a hybrid architecture that consist of mixed connections among P2P and Cloud nodes that can interoperate together. We have chosen to represent the P2P nodes as Planet Lab machines over the world and the cloud nodes using a Cloud provider's network. First we present an experimental validation of the Cloud infrastructure's ability to distribute streaming sessions with respect to some key streaming QoS parameters: jitter, throughput and packet losses. Next we show the results obtained from different test scenarios, when a hybrid distribution network is used. The scenarios measure the improvement of the multimedia QoS parameters, when nodes in the streaming distribution network (located in different continents) are gradually moved into the Cloud provider infrastructure. The overall conclusion is that the QoS of a streaming service can be efficiently improved, unlike in traditional P2P systems and CDN, by deploying a hybrid streaming architecture. This enhancement can be obtained by strategic placing of certain distribution network nodes into the Cloud provider infrastructure, taking advantage of the reduced packet loss and low latency that exists among its datacenters.

The problem of fairly distributing the capacity of a network among a set of sessions has been widely studied. In this problem, each session connects via a single path a source and a destination, and its goal is to maximize its assigned transmission rate (i.e., its throughput). Since the links of the network have limited bandwidths, some criterion has to be defined to fairly distribute their capacity among the sessions. A popular criterion is max-min fairness that, in short, guarantees that each session i gets a rate λisuch that no session s can increase λswithout causing another session s' to end up with a rate λs/s. Many max-min fair algorithms have been proposed, both centralized and distributed. However, to our knowledge, all proposed distributed algorithms require control data being continuously transmitted to recompute the max-min fair rates when needed (because none of them has mechanisms to detect convergence to the max-min fair rates). In this paper we propose B-Neck, a distributed max-min fair algorithm that is also quiescent. This means that, in absence of changes (i.e., session arrivals or departures), once the max min rates have been computed, B-Neck stops generating network traffic. Quiescence is a key design concept of B-Neck, because B-Neck routers are capable of detecting and notifying changes in the convergence conditions of max-min fair rates. As far as we know, B-Neck is the first distributed max-min fair algorithm that does not require a continuous injection of control traffic to compute the rates. The correctness of B-Neck is formally proved, and extensive simulations are conducted. In them, it is shown that B-Neck converges relatively fast and behaves nicely in presence of sessions arriving and departing.

In this brief announcement we propose B-Neck, a max-min fair distributed algorithm that is also quiescent. As far as we know, B-Neck is the first max-min fair distributed algorithm that does not require a continuous injection of control traffic to compute the rates. When changes occur, affected sessions are asynchronously informed, so they can start the process of computing their new rate (i.e., sessions do not need to poll the network for changes). The correctness of B-Neck is formally proved, and extensive simulations are conducted. In them it is shown that B-Neck converges relatively fast and behaves nicely in presence of sessions arriving and departing.

Since its appearance, peer-to-peer technology has given raise to various multimedia streaming applications. Today, cloud computing offers different service models as a base for successful end user applications. In this paper we propose joining peer-to-peer and cloud computing into new architectural realization of a distributed cloud computing network for multimedia streaming, in a both centralized and peer-to-peer distributed manner. This architecture merges private and public clouds and it is intended for a commercial use, but in the same time scalable to offer the possibility of non-profitable use. In order to take advantage of the cloud paradigm and make multimedia streaming more efficient, we introduce APIs in the cloud, containing build-in functions for automatic QoS calculation, which permits negotiating QoS parameters such as bandwidth, jitter and latency, among a cloud service provider and its potential clients.

In this paper we generalize the Continuous Adversarial Queuing Theory (CAQT) model [5] by considering the possibility that the router clocks in the network are not synchronized. Clearly, this new extension to the model only affects those scheduling policies that use some form of timing. First, if all clocks run at the same speed, maintaining constant differences, we show that all universally stable policies in CAQT that use the injection time and the remaining path to schedule packets remain universally stable. These policies include, for instance, Shortest in System (SIS) and Longest in System (LIS). Then, if clock differences can vary over time, but difference is bounded, we show the universal stability of SIS and a family of policies related to LIS. The bounds we obtain in this case depend on the maximum difference between clocks. We then present a new policy that we call Longest in Queues (LIQ), which gives priority to the packet that has been waiting the longest in edge queues. This policy is universally stable and, if clocks maintain constant differences, the bounds do not depend on them. To finish, we provide with simulation results that compare the behavior of some of these protocols in a network with stochastic injection of packets.

Social applications associate a set of user defined keywords named tags when publishing social objects in order to locate them later. We propose T-DHT, a hybrid unstructured-structured DHT based approach, to cope with the high demanding requirements of social applications, in a fully scalable, distributed and balanced way. T-DHT behaves as a structured DHT when publishing "tag, social-object" associations, and as an unstructured filter driven network when searching for the social-objects by means of any tag combination. The publishing process stores tag information across node links in order to drive adequately the later search operations in at most O(Log(N)). The search process takes at most O(logN) node hops for any tag combination and uses the previously stored node link tag information. Although T-DHT has been devised to build a scalable social application infrastructure, it can also be applied to solve general peer-to-peer keyword search problems