Federated Learning (FL) enables collaborative training without centralizing data, essential for privacy compliance in real-world scenarios involving sensitive visual information.Most FL approaches rely on expensive, iterative deep network optimization, which still risks privacy via shared gradients.In this work, we propose FedHENet, extending the FedHEONN framework to image classification.By using a fixed, pretrained feature extractor and learning only a single output layer, we avoid costly local fine-tuning.This layer is learned by analytically aggregating client knowledge in a single round of communication using homomorphic encryption (HE).Experiments show that FedHENet achieves competitive accuracy compared to iterative FL baselines while demonstrating superior stability performance and up to 70% better energy efficiency.Crucially, our method is hyperparameter-free, removing the carbon footprint associated with hyperparameter tuning in standard FL.

In this work, we propose a Federated Learning (FL) method combining an ensemble of one-layer neural networks whose optimal parameters can be obtained through a non-iterative procedure. Therefore, unlike most state-of-the-art methods, the collaborative global model can be obtained using a single round of communication between all clients of the federated scheme. It presents a computationally efficient and incremental batch aggregation process that suits the needs of a realistic federated scenario, simplifying the management of the federated training process. The model provides the same performance in identically and non-identically distributed data scenarios. Besides, the model implements a Fully Homomorphic Encryption (FHE) scheme to enhance robustness against privacy leaks or attacks, enabling clients to offload computational work to the coordinator, which operates entirely on encrypted data. We achieve an efficient and secure distributed model with an improved representation capacity for this type of architecture. The source code used in the study is made publicly available.

Population growth has increased the demand for meat and dairy products, making livestock, especially cattle, key to meeting this demand. This has led to an increase in herd size, complicating efficient herd management. To meet this challenge, innovative technologies, such as monitoring collars, have been developed to improve individual animal management. This research work evaluates and compares three unsupervised anomaly detection methods to identify estrus in dairy cows from intensive farms, based on daily activity data recorded by a commercial monitoring collar. Data from two different dairy farms have been used and the results have been compared by evaluating the behavior both individually and at herd level. The results obtained show a good performance of the selected techniques in the individual animal models. Thus, this research demonstrates that these techniques can be very useful tools in farm management, providing valuable information, improving productivity and, consequently, increasing the economic performance of the farm.

In the current landscape, user opinions exert an unprecedented influence on the trajectory of companies. In the field of online review platforms, these opinions, transmitted through text reviews and numerical ratings, significantly shape the credibility of products and services. For this reason, detecting inappropriate reviews becomes crucial. This paper addresses the problem of automatic anomalous review detection using a novel approach based on Anomaly Detection in the field of Natural Language Processing (NLP). Unlike other NLP tasks, anomaly detection in texts is a relatively emerging area. In this paper, we present a pipeline for opinion filtering that poses the problem of discerning between normal opinions containing relevant information about an item and anomalous opinions with unrelated content. Its key functionalities include: Classifying the reviews, assigning normality scores, and generating explanations for each classification, indispensable for the human who normally moderates these platforms. To evaluate the model, several Amazon datasets were used to demonstrate that the performance obtained is robust, obtaining an average F1 score of 91.4 detecting anomalies in the most complex scenario. In addition, a comparative study of three explainability techniques was conducted with 241 participants to measure the impact on understanding the classifications of the model and to rank their perceived usefulness of explanations. As a result, we obtained a system with great potential to automate tasks related to online review platforms, offering insights into anomaly detection applications in textual data and showing the difficulties that arise when the task to be explained presents a subjectivity component.

Abstract One of the most common attacks is man‐in‐the‐middle (MitM) which, due to its complex behaviour, is difficult to detect by traditional cyber‐attack detection systems. MitM attacks on internet of things systems take advantage of special features of the protocols and cause system disruptions, making them invisible to legitimate elements. In this work, an intrusion detection system (IDS), where intelligent models can be deployed, is the approach to detect this type of attack considering network alterations. Therefore, this paper presents a novel method to develop the intelligent model used by the IDS, being this method based on a hybrid process. The first stage of the process implements a feature extraction method, while the second one applies different supervised classification techniques, both over a message queuing telemetry transport (MQTT) dataset compiled by authors in previous works. The contribution shows excellent performance for any compared classification methods. Likewise, the best results are obtained using the method with the highest computational cost. Thanks to this, a functional IDS will be able to prevent MQTT attacks.

Autoencoder networks are widely used in anomaly detection, however, their training can be computationally expensive, limiting their use in Edge Computing and Federated Learning scenarios, where devices are usually not very powerful. In addition, there are several ways to directly or indirectly attack the privacy of the data used by these networks, which is unacceptable in this type of scenario. Unlike traditional autoencoder networks, Deep AutoEncoder for Federated learning (DAEF) does not require several rounds of learning since it is a non-iterative method. This implies greater speed, less network traffic, and lower energy consumption, as well as preventing the privacy attacks common in iterative networks. In this paper, we present an architecture designed for the use of the DAEF network in Edge Computing and Federated Learning scenarios. Unlike other collaborative machine learning approaches, it is not server based. Consequently, all the stages of the learning process, including the model aggregation, are performed on the edge devices (nodes). Each edge node trains its local DAEF network asynchronously concerning the rest. Nodes that have completed their training can voluntarily request to add their local model information to the global one. In this way, there is not a unique aggregator node, but each node in the network will be in charge of adding its local learning to the global model. The federated learning management is handled by a coordinator node using a Message Queuing Telemetry Transport (MQTT) communication protocol, which can be assumed by any device in the network in case of failures, and the information exchanged between the nodes does not compromise the privacy of the original local datasets.

Federated learning (FL) is a distributed approach to developing collaborative learning models from decentralized data. This is relevant to many real applications, such as in the field of the Internet of Things, since the models can be used in edge computing devices. FL approaches are motivated by and designed to protect privacy, a highly relevant issue given current data protection regulations. Although FL methods are privacy-preserving by design, recently published papers show that privacy leaks do occur, caused by attacks designed to extract private data from information interchanged during learning. In this work, we present an FL method based on a neural network without hidden layers that incorporates homomorphic encryption (HE) to enhance robustness against the above-mentioned attacks. Unlike traditional FL methods that require multiple rounds of training for convergence, our method obtains the collaborative global model in a single training round, yielding an effective and efficient model that simplifies management of the FL training process. In addition, since our method includes HE, it is also robust against model inversion attacks. In experiments with big data sets and a large number of clients in a federated scenario, we demonstrate that use of HE does not affect the accuracy of the model, whose results are competitive with state-of-the-art machine learning models. We also show that behavior in terms of accuracy is the same for identically and non-identically distributed data scenarios.

This paper presents a novel, fast and privacy preserving implementation of deep autoencoders. DAEF (Deep AutoEncoder for Federated learning), unlike traditional neural networks, trains a deep autoencoder network in a non-iterative way, which drastically reduces training time. Training can be performed incrementally, in parallel and distributed and, thanks to its mathematical formulation, the information to be exchanged does not endanger the privacy of the training data. The method has been evaluated and compared with other state-of-the-art autoencoders, showing interesting results in terms of accuracy, speed and use of available resources. This makes DAEF a valid method for edge computing and federated learning, in addition to other classic machine learning scenarios.

In this study, we analyze the capability of several state of the art machine learning methods to predict whether patients diagnosed with CoVid-19 (CoronaVirus disease 2019) will need different levels of hospital care assistance (regular hospital admission or intensive care unit admission), during the course of their illness, using only demographic and clinical data. For this research, a data set of 10,454 patients from 14 hospitals in Galicia (Spain) was used. Each patient is characterized by 833 variables, two of which are age and gender and the other are records of diseases or conditions in their medical history. In addition, for each patient, his/her history of hospital or intensive care unit (ICU) admissions due to CoVid-19 is available. This clinical history will serve to label each patient and thus being able to assess the predictions of the model. Our aim is to identify which model delivers the best accuracies for both hospital and ICU admissions only using demographic variables and some structured clinical data, as well as identifying which of those are more relevant in both cases. The results obtained in the experimental study show that the best models are those based on oversampling as a preprocessing phase to balance the distribution of classes. Using these models and all the available features, we achieved an area under the curve (AUC) of 76.1% and 80.4% for predicting the need of hospital and ICU admissions, respectively. Furthermore, feature selection and oversampling techniques were applied and it has been experimentally verified that the relevant variables for the classification are age and gender, since only using these two features the performance of the models is not degraded for the two mentioned prediction problems.

Spectral clustering is a clustering paradigm that has been shown to be more effective in finding clusters with non-convex shapes than some traditional algorithms such as k-means.However, this algorithm is not directly applicable when the data is naturally distributed in different locations, as it happens in many Internet of Things scenarios.In this work, we propose a distributed spectral clustering to create a cooperative federated model to deal with those cases in which the data is distributed in different sites and with data privacy concerns.We demonstrate that sharing a minimal amount of information allows this distributed version of the spectral clustering to achieve good behavior for clustering several synthetic data sets.

One-class classification has gained interest as a solution to certain kinds of problems typical in a wide variety of real environments like anomaly or novelty detection. Autoencoder is the type of neural network that has been widely applied in these one-class problems. In the Big Data era, new challenges have arisen, mainly related with the data volume. Another main concern derives from Privacy issues when data is distributed and cannot be shared among locations. These two conditions make many of the classic and brilliant methods not applicable. In this paper, we present distributed singular value decomposition (DSVD-autoencoder), a method for autoencoders that allows learning in distributed scenarios without sharing raw data. Additionally, to guarantee privacy, it is noniterative and hyperparameter-free, two interesting characteristics when dealing with Big Data. In comparison with the state of the art, results demonstrate that DSVD-autoencoder provides a highly competitive solution to deal with very large data sets by reducing training from several hours to seconds while maintaining good accuracy.

Due to the frequent use of anomaly detection systems in monitoring and the lack of methods capable of learning in real time, this research presents a new method that provides such online adaptability. The method developed is called OSHULL (Online and Subdivisible Distributed Scaled Convex Hull) and bases its operation on the properties of scaled convex hulls. It begins building a convex hull, using a minimum set of data, that is adapted and subdivided along time to accurately fit the boundary of the normal class data. The method has been evaluated and compared to several main algorithms of the field using some real and artificial data sets. As a consequence, an algorithm has been obtained with online learning ability and easily configurable, all without diminishing its effectiveness in relation to other batch state-of-the-art methods. Finally, its execution can be carried out in a distributed and parallel way, which is an interesting advantage in the treatment of big data sets.

In this paper, a new one-class classification algorithm capable of working in distributed environments is presented. In it, convex hull is used to build the boundary of the target class defining the one-class problem in each of the distributed nodes. Therefore, we will consider several classifiers, each one determined using a given local data partition, and the goal is to obtain a global classification decision. In order to obtain this final decision, two different algebraic combination rules were proposed: 1) sum and 2) majority voting. Experimental results show that this method opens the possibility of tackling practical one-class classification problems in distributed big data scenarios in an efficient and accurate way.

In the scope of data analytics, the volume of a data set can be defined as a product of instance size and dimensionality of the data. In many real problems, data sets are mainly large only on one of these aspects. Machine learning methods proposed in the literature are able to efficiently learn in only one of these two situations, when the number of variables is much greater than instances or vice versa. However, there is no proposal allowing to efficiently handle either circumstances in a large-scale scenario. In this brief, we present an approach to integrally address both situations, large dimensionality or large instance size, by using a singular value decomposition (SVD) within a learning algorithm for one-layer feedforward neural network. As a result, a noniterative solution is obtained, where the weights can be calculated in a closed-form manner, thereby avoiding low convergence rate and also hyperparameter tuning. The proposed learning method, LANN-SVD in short, presents a good computational efficiency for large-scale data analytic. Comprehensive comparisons were conducted to assess LANN-SVD against other state-of-the-art algorithms. The results of this brief exhibited the superior efficiency of the proposed method in any circumstance.

Microarray data classification is a difficult problem for computational techniques due to its inherent properties mainly, its imbalanced distribution and small sample size. Machine learning has been widely employed for handling this type of data predominantly applying two-class classification techniques. However, one-class approach has the ability to deal with imbalanced distribution and unexpected noise in the data. To deal with these situations it is considered that the best option is using the minority class as the target concept. This is reinforced by the idea of obtaining a classifier able to adjust itself to the specificity of the given class despite sacrificing the additional information about the second class. Although this consideration appears in different research, there are no thorough studies that prove it experimentally. In this paper, we investigate the suitability of employing minority class as the concept target in one-class classification to handle the class imbalance problem. A study over several microarray data sets is included. The results confirm that the use of minority class allows us to obtain better performance in one-class classification.

This paper presents a novel distributed one-class classification approach based on an extension of the ν-SVM method, thus permitting its application to Big Data data sets. In our method we will consider several one-class classifiers, each one determined using a given local data partition on a processor, and the goal is to find a global model. The cornerstone of this method is the novel mathematical formulation that makes the optimization problem separable whilst avoiding some data points considered as outliers in the final solution. This is particularly interesting and important because the decision region generated by the method will be unaffected by the position of the outliers and the form of the data will fit more precisely. Another interesting property is that, although built in parallel, the classifiers exchange data during learning in order to improve their individual specialization. Experimental results using different datasets demonstrate the good performance in accuracy of the decision regions of the proposed method in comparison with other well-known classifiers while saving training time due to its distributed nature.

The LOw Carbon At Work (LOCAW) project has studied the everyday behavior of employees in six organizations in order to achieve a more sustainable Europe. Of these six, four organizations were involved in backcasting workshops to obtain future scenarios aimed at significantly improving engagement with pro-environmental behaviors by 2050. From these scenarios policies were extracted from the workshop participants that achieve this aim in their organization. Agent Based Models (ABM) were designed to model the organizations using actual information from the organization; the design also placed special emphasis on the representation of the social network. ABMs were then used to simulate the effects of the different policies derived from the backcasting scenarios. In this paper, the results for two organizations, UDC and Aquatim, are shown. These experimental results demonstrate the influence of different social networks and internal motivations of employees to determine the effectiveness of a given policy.

Many real problems in machine learning are of dynamic nature. In that cases, the model used for the learning process should work in real time and have the ability to act and react by itself, adjusting its controlling parameters, even its structures, depending on the requirements of the process. In a previous work, the authors proposed an online learning method for two-layer feedforward neural networks that presents two main characteristics. Firstly, it is effective in dynamic environments as well as in stationary contexts. Secondly, it allows to incorporate new hidden neurons during learning without loosing the knowledge already acquired. In this paper, we extended this previous algorithm including a mechanism to automatically adapt the network topology according with the needs of the learning process. This automatic estimation technique is based on the Vapnik-Chervonenkis dimension. The theoretical basis for the method is given and its performance is illustrated by means of its application to different system identification problems. The results confirm that the proposed method is able to check whether new hidden units should be added depending on the requirements of the online learning process.

A technique for adjusting a minimum volume set of covering ellipsoids technique is elaborated. Solutions to this problem have potential application in one-class classification and clustering problems. Its main original features are: 1) It avoids the direct evaluation of determinants by using diagonalization properties of the involved matrices, 2) it identifies and removes outliers from the estimation process, 3) it avoids binary variables resulting from the combinatorial character of the assignment problem that are replaced by continuous variables in the range [0,1], 4) the problem can be solved by a bilevel algorithm that in its first level determines the ellipsoids and in its second level reassigns the data points to ellipsoids and identifies outliers based on an algorithm that forces the Karush-Kuhn-Tucker conditions to be satisfied. Two theorems provide rigorous bases for the proposed methods. Finally, a set of examples of application in different fields is given to illustrate the power of the method and its practical performance.

Vibration analysis is one of the most used techniques for predictive maintenance in high-speed rotating machinery. Using the information contained in the vibration signals, a system for alarm detection and diagnosis of failures in mechanical components of power wind mills is devised. As previous failure data collection is unfeasible in real life scenarios, the method to be employed should be capable of discerning between failure and normal data, being only trained with the latter type. Other interesting capability of such a method is the possibility of measuring the evolution of the failure. Taking into account these restrictions, a method that uses the one-class-ν-SVM paradigm is employed. In order to test its adequacy, three different scenarios are tested: (a) a simulated scenario, (b) a controlled experimental scenario with real vibrational data, and (c) a real scenario using vibrational data captured from a windmill power machine installed in a wind farm in North West Spain. The results showed not only the capabilities of the method for detecting the failure in advance to the breakpoint of the component in all three scenarios, but also its capacity to present a qualitative indication on the evolution of the defect. Finally, the results of the SVM paradigm are compared to one of the most used novelty detection methods, obtaining more accurate results under noisy circumstances.

Many real scenarios in machine learning are non-stationary. These challenges forces to develop new algorithms that are able to deal with changes in the underlying problem to be learnt. These changes can be gradual or abrupt. As the dynamics of the changes can be different, the existing machine learning algorithms exhibit difficulties to cope with them. In this work we propose a new method, that is based in the introduction of a forgetting function in an incremental online learning algorithm for two-layer feedforward neural networks. This forgetting function gives a monotonically crescent importance to new data. Due to this fact, the network forgets in presence of changes while maintaining a stable behavior when the context is stationary. The theoretical basis for the method is given and its performance is illustrated by evaluating its behavior. The results confirm that the proposed method is able to work in evolving environments.

In this paper, a novel supervised architecture for binary classification based on local modelling and information theory is described. The architecture is composed of two steps: in the first one, a separating borderline between the two classes is piecewise constructed by a set of centroids calculated by a modified clustering algorithm, based on information theory; each of these centroids define a region where, in the second step of the proposed architecture, a hyperplane is constructed and adjusted by means of one-layer neural networks. This new method allows for binary classification while maintaining adequate use of computational resources, a common problem for machine learning methods. The proposed architecture is applied over classical benchmark classification problems and data sets, and its results are compared with those obtained by other well-known statistical and machine learning classifiers.

In this paper, a new pattern recognition method is presented. It is based on the combination of two techniques. The first is a modified version of growing neural gas, and the second is a set of proximal support vector machines. The aim of the former is to obtain a topology in the network that defines different local regions into the input space; and the goal of the latter is to fit a set of local classifiers for each one of the regions. The efficiency of the algorithm is validated on two data sets and is compared to another standard algorithm. The results obtained by the method presented exhibit a good performance in all cases.

This paper introduces a learning method for two-layer feedforward neural networks based on sensitivity analysis, which uses a linear training algorithm for each of the two layers. First, random values are assigned to the outputs of the first layer; later, these initial values are updated based on sensitivity formulas, which use the weights in each of the layers; the process is repeated until convergence. Since these weights are learnt solving a linear system of equations, there is an important saving in computational time. The method also gives the local sensitivities of the least square errors with respect to input and output data, with no extra computational cost, because the necessary information becomes available without extra calculations. This method, called the Sensitivity-Based Linear Learning Method, can also be used to provide an initial set of weights, which significantly improves the behavior of other learning algorithms. The theoretical basis for the method is given and its performance is illustrated by its application to several examples in which it is compared with several learning algorithms and well known data sets. The results have shown a learning speed generally faster than other existing methods. In addition, it can be used as an initialization tool for other well known methods with significant improvements.

Training multilayer neural networks is typically carried out using descent techniques such as the gradient-based backpropagation (BP) of error or the quasi-Newton approaches including the Levenberg-Marquardt algorithm. This is basically due to the fact that there are no analytical methods to find the optimal weights, so iterative local or global optimization techniques are necessary. The success of iterative optimization procedures is strictly dependent on the initial conditions, therefore, in this paper, we devise a principled novel method of backpropagating the desired response through the layers of a multilayer perceptron (MLP), which enables us to accurately initialize these neural networks in the minimum mean-square-error sense, using the analytic linear least squares solution. The generated solution can be used as an initial condition to standard iterative optimization algorithms. However, simulations demonstrate that in most cases, the performance achieved through the proposed initialization scheme leaves little room for further improvement in the mean-square-error (MSE) over the training set. In addition, the performance of the network optimized with the proposed approach also generalizes well to testing data. A rigorous derivation of the initialization algorithm is presented and its high performance is verified with a number of benchmark training problems including chaotic time-series prediction, classification, and nonlinear system identification with MLPs.

Proper initialization of neural networks is critical for a successful training of its weights. Many methods have been proposed to achieve this, including heuristic least squares approaches. In this paper, inspired by these previous attempts to train (or initialize) neural networks, we formulate a mathematically sound algorithm based on backpropagating the desired output through the layers of a multilayer perceptron. The approach is accurate up to local first order approximations of the nonlinearities. It is shown to provide successful weight initialization for many data sets by Monte Carlo experiments.

The article presents a method for learning the weights in one-layer feedforward neural networks minimizing either the sum of squared errors or the maximum absolute error, measured in the input scale. This leads to the existence of a global optimum that can be easily obtained solving linear systems of equations or linear programming problems, using much less computational power than the one associated with the standard methods. Another version of the method allows computing a large set of estimates for the weights, providing robust, mean or median, estimates for them, and the associated standard errors, which give a good measure for the quality of the fit. Later, the standard one-layer neural network algorithms are improved by learning the neural functions instead of assuming them known. A set of examples of applications is used to illustrate the methods. Finally, a comparison with other high-performance learning algorithms shows that the proposed methods are at least 10 times faster than the fastest standard algorithm used in the comparison.

The recognition of accelerative and decelerative patterns in the fetal heart rate (FHR) is one of the tasks carried out manually by obstetricians when they analyze cardiotocograms for information respecting the fetal state. An approach based on artificial neural networks formed by a multilayer perceptron (MLP) is developed. However, since the system utilizes the FHR signal as direct input, an anterior stage must be incorporated that applies a principal component analysis (PCA) so as to make the system independent of the signal baseline. Furthermore, the introduction of multiresolution into the PCA has resolved other problems that were detected in the application of the system. Presented in this paper are the results of validation of these systems designated the PCA-MLP and multiresolutlon principal component analysis (MR-PCA) systems against three clinical experts.