Federated Learning (FL), an emerging decentralized Machine Learning (ML) approach, offers an effective avenue for training models on distributed data while safeguarding consumer privacy. Nevertheless, adversarial attacks initiated by malicious clients gaining access to confidential consumer data may deteriorate the learning process and performance. We develop the novel MADE-PI technique for the conventional FL that identifies and excludes malicious clients in the early model training phases. Our method incurs a novel local models’ preprocessing step performed before the aggregation procedure, in which we analyze the distribution of PI scores assigned to each client and exclude potentially poisoned data sources. We introduce Proportional – Integral (PI) scores as distance-based detection metrics inspired by PID control that consider the current distance between clients and the history of distances, through the proportional and integral terms respectively. Our empirical study on three distinct consumer application FL datasets, which represent use-cases from the intelligent transportation, handwriting recognition and medical scenarios, verifies that our approach is successful in identification and removal of compromised clients, mitigating threats to consumer model integrity and the application performance. The possible gain in learning effectiveness and efficiency against conventional methods is evaluated and analyzed.

In this research, we address a critical vulnerability of Federated Learning (FL) in practical applications: their susceptibility to adversarial attacks from malicious participants, which can severely degrade a global model’s performance. We propose MADE-PI, a novel defense mechanism designed to identify and exclude malicious clients early in the training process. Our method introduces a pre-aggregation step where the server evaluates each client update using a Proportional-Integral (PI) score. This score is a unique distance-based metric inspired by the PID controller, where a proportional term P measures the immediate deviation of a client’s update, and an integral term I tracks its behavior over time. By analyzing the distribution of these scores, our technique effectively flags and removes malicious contributions. We validate our approach through an empirical study on three datasets representing various applications of FL, including intelligent transportation, medical image analysis, and handwriting recognition. We concentrate on data poisoning as a highly relevant class of attacks. The results demonstrate that MADE-PI outperforms conventional defenses by providing superior malicious client detection and improving both learning efficiency and final model accuracy.

Machine Learning (ML) systems are integrated with other parts of modern cyberinfrastructure (CI), which forms the novel ML with Integrated Network (MLIN) structure. Various security tools are already employed in CI to detect malicious attacks. However, their operation is limited to a certain CI part or MLIN component without considering their integration and the ML end performance as the major indicator. We design and implement Knowledge Integration System Service (KISS) that introduces a novel approach to detect and classify attacks into adversarial attacks against ML end system (A-Attacks), and against base MLIN infrastructure (B-Attacks). Unlike traditional security mechanisms, KISS examines how attacks on individual MLIN components impact the overall ML end system performance, and extracts and integrates knowledge from each MLIN CI component to better distinguish between the attack types. As our major contributions, we (1) investigate the effects of A- and B-Attacks on ML systems, actively used in industry. We (2) develop the KISS prototype and verify it in practice. We demonstrate how KISS can be integrated into already established MLIN CI and combined with the traditional security tools. Our experiments verify that KISS improves attack detection and their initial classification between A- and B-Attacks in MLIN operational setups.

In this paper, we present an innovative approach to automate key event detection and highlights generation from soccer match videostreams that allows to improve accuracy and reliability, as well as to reduce data consumption and training time. Our method segments the videostream into distinct frames based on camera angles and activities, and integrates intelligent video analytics with additional visual information provided by broadcasters. As our major novelty in comparison to other intelligent soccer video analysis approaches, we deploy a Multi-Class Image Classifier to segment the video into wide-angle overviews, close-ups, and in-game replays, which allows us to improve the event detection performance and the quality of generated highlights. As another major novelty, we leverage YOLOv8 to detect events such as bookings, substitutions, and goals based on the additional information portrayed by broadcasters during the game. We evaluate our approach and demonstrate its advantage, when the additional information from broadcasters is available, against existing ones that analyze only the actions happening in the scene itself, such as the players' current positions and their actions between the frames. We evaluate our pipeline on a real soccer game recording and compare the highlights it generates with the official highlights provided by the broadcaster. Our pipeline demonstrates ample performance and is able to detect all key events in the game.

In this paper, we investigate the robustness of the ML end application performance to network Quality of Service (QoS) degradation, and the ways to improve it. We introduce a novel system approach to define the Machine Learning (ML) with Integrated Networks (MLINs) and describe how ML end performance can be employed to adjust network hyperparameters in order to prevent system Data Quality decrease. We investigate the interrelations between the network QoS degradation during the data transmission and ML image classification performance. We demonstrate how the studied interrelationships can be employed to produce recommendations on network adjustment in order to improve MLIN robustness. In particular, we propose an example of MLIN feedback system design that employs ML end performance as the major indicator to produce recommendations on network hyperparameters adjustment aimed at improving MLIN robustness.

The evolution of the Machine Learning (ML) has led to the emergence of Transfer Learning (TL) approach, which allows reusing pretrained industrial ML systems after their input data values or even an application domain has changed. In this paper, we investigate the TL process and its impact on the performance of ML-end systems with integrated network facilities. Especially, we focus on ML-systems designed for the classification of image media-files, transmitted over a network. Packet loss in a network is considered as a major input Data Quality (DQ) deterioration factor that can result in ML system classification performance degradation after pretraining on good inputs. To investigate the typical industrial TL process, we study the relationships between the ML model's last layer weights, hyperparameters, and classification performance throughout the retraining process. In addition, we conduct an empirical study to evaluate how the TL affects ML model performance in real application scenarios. For our experiments, we employ real image media-files, and transmit them over a real wireless network with inherent data losses for a classification on a remote ML-end system. According to our results, retraining ML models on corrupted data allows to enhance their robustness to a DQ degradation in the considered image classification scenarios. However, DQ influence on the ML system performance may vary depending on the data and system types.

The paper presents a design and an implementation of an intelligent detector of a novel "colluded applications" attack on user's privacy in Android OS devices, which employs recurrent neural network (RNN) models. The paper reports the results of an empirical study that involved the attack research, data collection and pre-processing, the choice of the RNN model for a detector design, multiple detector implementations, their performance evaluation and analysis, and finally, an Android app realization and execution on a real device. We investigate and analyze multiple attack scenarios and the attack influence on such technological signals as memory consumption and a CPU's cores clock speed. For the attack detection, a few detectors exploring multiple RNN models are designed, implemented, and examined. The detectors employ various RNN models, such as a simple recurrent neural network, a long short-term memory, and a gated recurrent unit. Each model's performance in detecting multiple attack scenarios is evaluated and analyzed in order to compare classification models against various criteria.

The paper investigates the feasibility of using machine learning (ML) applications for an anomaly-based unknown attack detection by monitoring system parameters on resource-constrained Android mobile devices. It introduces colluded applications example as a use case of unknown sophisticated attacks, examines its formalization models and possible scenarios as well as develops the machine learning classifiers for its detection. We investigate multiple attack cases and record them to produce datasets characterizing a memory consumption and a CPU cores clock speed changes during the attacks that could be employed to design classifiers based on ML techniques. For the attack detection, a few ML classifiers are presented and examined. The classifiers employ various ML techniques and models, such as feed-forward neural network and long short-term memory. Model performance in detecting multiple attack scenarios are presented. Classification models compared against various criteria.

The growing number of fitness trackers and dedicated smartphone applications indicate the importance of human activity recognition tasks in everyday life. Wearable devices tend to incorporate a broad spectrum of various sensors, such as an accelerometer, a gyroscope, a heart rate sensor, etc. It is essential to understand what type of data and algorithm contribute to a better overall activity recognition rate, especially for real-time activity recognition based on real-time data streams. In this paper, we focus on comparing and contrasting different machine learning techniques and gauge their performance in terms of human activity recognition accuracy. We investigate and analyze sensor mounting position, sensor types, and data type influence on activity recognition accuracy. Also, we render and analyze the benefits of using a RNN classifier over a more static classifier such as J48 and Naive Bayes algorithms in our empirical study.

This demonstration presents an application for a system security score evaluation. The security score represents overall security level of the Android OS based device. The proposed application is based on the library that uses standard Android OS API and Google SafetyNet library. We followed up modular approach that allows to use our library independently from our application in any stock Android OS version 5.0 and above. Application evaluates overall security level, gives an explanation of the each parameter, gives advices how to improve overall security of the device, and allows to simulate the evaluation that shows how each parameter affects the security level.

The paper aims at developing recommendations on the design and use of human activity recognition systems based on wearable sensors and their applications. Both the system structure and the application design challenges are addressed. Various design problems are investigated such as the selection of a number and types of sensors, amount of samples used for recognition, choice of machine learning techniques and the features set for a classifier design as well as the type of a training model employed. The design is implemented on the Android smartphone sensors platform and tested by conducting an empirical study in a real-life environment. The study results are provided and discussed. The recommendations on the sensor systems design are derived.

Computational Intelligence (CI) algorithms/techniques are packaged in a variety of disparate frameworks/applications that all vary with respect to specific supported functionality and implementation decisions that drastically change performance. Developers looking to employ different CI techniques are faced with a series of trade-offs in selecting the appropriate library/framework. These include resource consumption, features, portability, interface complexity, ease of parallelization, etc. Considerations such as language compatibility and familiarity with a particular library make the choice of libraries even more difficult. The paper introduces MeetCI, an open source software framework for computational intelligence software design automation that facilitates the application design decisions and their software implementation process. MeetCI abstracts away specific framework details of CI techniques designed within a variety of libraries. This allows CI users to benefit from a variety of current frameworks without investigating the nuances of each library/framework. Using an XML file, developed in accordance with the specifications, the user can design a CI application generically, and utilize various CI software without having to redesign their entire technology stack. Switching between libraries in MeetCI is trivial and accessing the right library to satisfy a user's goals can be done easily and effectively. The paper discusses the framework's use in design of various applications. The design process is illustrated with four different examples from expert systems and machine learning domains, including the development of an expert system for security evaluation, two classification problems and a prediction problem with recurrent neural networks.

Computational intelligence techniques are attracting more and more attention in NLP and text analysis applications. This paper is devoted to their use in automatic question generation based on text analysis with the goal to develop the computational intelligence framework that should automate or semi-automate the process of quiz and exam question generation. The framework operation is based on information retrieval and NLP algorithms. It incorporates the application of production rules, LSTM neural network models, and other intelligent techniques. It allows generating multiple choice questions, true and false questions as well as "Wh"-type (What? When? How?) questions. Automation procedures for each type question generation are developed, presented and analyzed. The typical challenges in framework development and application are considered and possible solutions are discussed. The results of the framework application and its use for quiz generation in a real college class are presented and analyzed.

This paper investigates the feasibility of employing artificial neural network techniques for solving fundamental cryptography problems, taking quadratic residue detection as an example. The problem of quadratic residue detection is one which is well known in both number theory and cryptography. While it garners less attention than problems such as factoring or discrete logarithms, it is similar in both difficulty and importance. No polynomial-time algorithm is currently known to the public by which the quadratic residue status of one number modulo another may be determined. This work leverages machine learning algorithms in an attempt to create a detector capable of solving instances of the problem more efficiently. A variety of neural networks, currently at the forefront of machine learning methodologies, were compared to see if any were capable of consistently outperforming random guessing as a mechanism for detection. Surprisingly, neural networks were repeatably able to achieve accuracies well in excess of random guessing on numbers up to 20 bits in length. Unfortunately, this performance was only achieved after a super-polynomial amount of network training, and therefore we do not believe that the system as implemented could scale to cryptographically relevant inputs of 500 to 1000 bits. This nonetheless suggests a new avenue of attack in the search for solutions to the quadratic residues problem, where future work focused on feature set refinement could potentially reveal the components necessary to construct a true closed-form solution.

As the amount of data available to researchers continues to grow, sharp class distinctions generated through Boolean logic fundamentally limit researchers' abilities to adequately process data into useable information. The application of fuzzy systems to real world problems allows us to synthesize this raw data into meaningful results. This paper demonstrates how fuzzy theory can be applied to real life application's design where rules can be generated explicitly through an expert system, or implicitly derived through a hybrid neural-fuzzy system. The paper illustrates the success of the fuzzy system use in applications by providing examples in the context of mobile device security evaluation and air quality evaluation and prediction. The examples represent important application areas within engineering and information technology domains.

This paper analyzes the advantages and disadvantages of two types of self-validating software that may run on into measurements devices: programs, which realize interval computations, and programs, which process inaccurate initial data that are expressed in terms of fuzzy intervals. The proposed analysis is presented from the viewpoint of metrology requirements. It is demonstrated that expressing measurement uncertainties in the form of fuzzy intervals results in the software that is better suited for its application in measurement systems.

This paper presents a feasibility study of novel attack detection mechanisms in wireless sensor networks (WSN) based on detecting anomalies and changes in sensor signals and data values. Typical WSN attacks are considered in the empirical study of various attack detection techniques utilizing features based on sensor signal strength and other WSN technological parameters and using machine learning classification techniques such as clustering, rule learners, and neural networks. For the attack detection implementation the study employed WSN built from Sun kits available on the market and extended Sensor Network Anomaly Detection System (SNADS) framework of methods and tools.

Data quality (DQ) is essential to achieve data trustworthiness, as it assures that data is free of errors, complete, and consistent. This paper proposes an approach to evaluate DQ in multichannel sensor networks and systems with heterogeneous data sources. The approach integrates various DQ indicators ranging from traditional data accuracy metrics to network security and business performance measures. It demonstrates the advantage of including security metrics into the DQ evaluation for the design optimization of data fusion procedures and even the whole data collection and communication systems. The DQ metrics composition and calculus are discussed. However, the major attention is paid to the analysis of the relationship between conventional data accuracy metrics and network security indicators.

Biologically-realistic multi-precision spiking neural network (SNN) simulation is designed and implemented on a new GPU device Radeon™ HD 7970 using OpenCL framework. The implementation aims to investigate the role of time precision in simulated SNNs. Simulation methods and GPU platforms are reviewed. Simulation model and process are presented and analyzed. The GPU model is capable of simulating a SNN with up to two million neurons. GPU and CPU results are directly verified and found to match exactly.

A novel GPU-based simulation of spiking neural networks is implemented as a hybrid system using Parker-Sochacki numerical integration method with adaptive order. Full single-precision floating-point accuracy for all model variables is achieved. The implementation is validated with exact matching of all neuron potential traces from GPU-based simulation versus those of a reference CPU-based simulation. A network of 4096 Izhikevich neurons simulated on an NVIDIA GTX260 device achieves real-time performance with a speedup of 9 compared to simulation executed on Opteron 285, 2.6-GHz device.

The paper puts forward a concept of cognitive sensor networks and investigates a feasibility of artificial neural networks application for its realization. It describes a design of novel hierarchical configurations imitating the structural topology of brain-like architectures. They are composed from artificial neural networks distributed over network platforms with limited resources. The paper examines a cognition idea based on its implementation through the signal change detection. The novel multilevel neural networks architectures are designed and tested in sensor networks built from Crossbow Inc. sensor kits. The results are compared against conventional multilayer perceptron structures in terms of both functional efficiency and resource consumption.

The paper investigates the feasibility of fuzzy models application in measurement procedures. It considers the problem of measurement information fusion from different sources, when one of the sources provides predictions regarding approximate values of the measured variables or their combinations. Typically, this information is given by an expert but may be mined from available data also. This information is formalized as fuzzy prediction models and is used in combination with the measurement results to improve the measurement accuracy. The properties of the modified estimates are studied in comparison with the conventional ones. The conditions when fuzzy models application can achieve a significant accuracy gain are derived, the gain value is evaluated, and the recommendations on fuzzy prediction model production and formalization in practical applications are given.

The paper describes development of the universal fuzzy basic expert system shell (BESS). The goal of the Bess system is to provide a fuzzy expert system development framework that is not only language independent, but is not attached to a specific expert system shell either. An expert system that is written in Bess may be written exactly once regardless of the number of shell implementations and modifications on which it is intended to run; the only changes required in the system are the additions of new data, or the modifications of existing data. This arrangement completely separates the knowledge and data bases in an expert system from the rules engine used to make inferences. Bess is built on top of a simple, easy to use XML compliant language for describing expert systems in terms that are commonly used for this purpose. This software is used for developing an expert system for computer security evaluation.

This paper documents the results of experimental simulations designed to compare the performance of Multilayer Perceptron (MLP) and Radial Basis Function (RBF) based sensor signal change detection systems. The two systems are simultaneously executed in parallel on the same input signals. Both systems share an identical implementation with the exception of the activation function used in the hidden layers of the artificial neural networks. Previous experiments have employed only Multilayer Perceptrons with sigmoidal activation functions. The results of these experiments quantitatively show the advantages and disadvantages of Radial Basis neural activation for both the function prediction and function correlation neural networks tested.

The paper introduces a modified time-based multilayer perceptron (MTBMLP), which is a complex structure composed by a few time-based multilayer perceptrons. This modification reduces connections, isolates information for each function and produces knowledge about the system of functions as a whole. This neural network is applied for novelty and change detection in signals delivered by sensor networks and for edge detection in image processing. In both applications a MTBMLP is utilized for function predictions and, after a further structure development is implemented, for an error prediction also. In sensor network applications, a number of experiments with Crossbow sensor kits and the MTBMLP acting as a function predictor have been conducted and analyzed for detecting a significant change in signals of various shapes and nature. A series of experiments with Lena image have been conducted for edge detection applications. The results demonstrate that MTBMLP is more efficient and reliable than other methodologies in sensor network change detection and that its application in change detection is more effective than in edge detection.

The paper develops intelligent protocols, enhancing efficiency, reliability and security based on detection of sensor signal changes, and describes their possible implementation on the lower level in sensor networks. Embedded into the communication protocols, the signal change detection will allow data compression for improving network efficiency. It might enhance reliability and security also. The protocol utilizes a neural network function prediction methodology in order to determine if the sensor signals have changed. In addition to the change detection system, a modification of a standard neural network function predictor is proposed that allows the detection system to learn quickly how to accurately predict next sensor outputs. The parameter choice and the relationship between the threshold values and false positive and negative rates are studied. The protocol is implemented and tested in real life environments with sensor networks built from Crossbow MICA-2 motes. Sensor Network Anomaly Detection Systems, which is designed to become a protocol's core utility, is described. The test results are analyzed and recommendations on applications are provided.

The paper considers the problem of improving accuracy and reliability of measurement information acquired by sensor networks. It offers the way of integrating sensor measurement results with association information available or a priori derived at aggregating nodes. The models applied for describing both sensor results and association information are reviewed with consideration given to both neuro-fuzzy and probabilistic models and methods. The information sources, typically available in sensor systems, are classified according to the model (fuzzy or probabilistic), which seems more feasible to be applied. The integration problem is formalized as an optimization problem.

The paper considers the problem of measurement information fusion from different sources, when one of the sources is an information about approximate values of the measured variables or their combinations. The information is given with fuzzy models and is used in combination with the measurement results. The properties of the modified estimates are studied in comparison with the conventional ones. The conditions when an expert's information application can give a high gain are derived, the gain value is estimated, the recommendations to an expert on making predictions are given. The possible gain in measurement result efficiency in geophysical applications is analyzed.

The paper investigates the method of a fuzzy system design through its approximation with neural networks. It concentrates on further simplification by replacement of a Gaussian radial basis function with its linear and piecewise linear approximation. Different approximating possibilities are tested on four controllers chosen as benchmarks.

The paper combines two major neuro-fuzzy applications in power engineering: stabilizing power systems at a generation stage and reducing disturbances at a delivery stage. It presents a neural-fuzzy classifier for recognition of power disturbances and a fuzzy excitation controller comprising both the exciter and the power system stabilizer.

The paper studies the problem of information uncertainty evaluation in modern engineering and technology applications, and especially the system design. It analyses the virtual environment design and engineering measurement. Information typical for those applications is classified according to its uncertainty types. Uncertainty sources are identified. Fuzzy theory models are proposed. Examples of their applications in characteristic problems are given.

This paper is devoted to the development of a fast industrial design of an embedded fuzzy controller for an unstable plant starting from an initial fuzzy system developed in conjunction with an expert. Through prototyping and implementation on a general-purpose microprocessor the plant can then be tested and used to iteratively identify and control the plant on the desired control system criteria is achieved. This paper describes the design, implementation and testing of an embedded fuzzy controller as a simple, practical solution to a regulation problem in reefer refrigeration systems. Fuzzy logic was utilized for the ease with which it enables expert opinion expressed within the framework of a natural language to be converted into a PID-like control algorithm. By using the Motorola HC12 MCU with its inbuilt fuzzy logic instruction set a rapid prototype design was implemented and the application of fuzzy logic and iterative identification and control to general industrial refrigeration systems was found to be a possible economically feasible solution.

Low cost microprocessors cannot always devote the resources necessary to compute a fuzzy system, and this can be a deterrent in its application. The purpose of this work is to demonstrate that neural networks are a viable form for implementing fuzzy systems in a practical cost effective application. A neural network can be trained to efficiently approximate a fuzzy control surface to a desired degree of accuracy. The paper proposes a neuro-fuzzy synergetic design procedure consisting of a fuzzy controller design and its implementation with a radial basis function neural network. The trade-offs associated with accuracy, speed and processing requirements are addressed, and the realization results are then presented and discussed.

We consider an intelligent measurement as a process of the result development based on all sorts of information which can be used with an application of intelligent technologies and methodologies. In real life systems, applied for control or testing, the measurement results compose only a part of information which is actually available and used for control purposes. An attempt to classify all the information available is given. As any measurement result incorporates some error, it can be considered as a fuzzy granule and a measurement process can be looked at as a fuzzy granulation process.

This paper attempts to analyse and classify different methods of fuzzy control (FC) design. It aims at providing some practical recommendations on how to design a FC and how to choose the right design method. The goal is to develop a more or less systematic procedure of a FC design which can be used as a reference. The design process can be roughly divided into two stages: an initial choice of the controller structures and parameters, and their further adjustment. Recommendations are given at the first stage regarding a choice of the structure, scaling factors, rules and membership functions, though the main attention is focused on the membership functions and scaling factors. Different methodologies, such as neural networks, genetic/evolutionary algorithms, are considered at the second stage.