ABSTRACT In recent years, the research area of cybersecurity has placed an emphasis on machine learning as an efficient security construct to be used in identifying the malignant malware within the contemporary anti‐malware programs. This study focuses on the crucial issue of privacy of data and ensuring that efficiency is achieved in detecting and classifying malware that is used. It is motivated by the fact that there is an increasing demand for privacy and secure machine learning methods in response to the growing threats to centralized data systems. The optimization‐based deep federated learning is an alternative approach to maintaining privacy without creating centralized points of data gathering. This paper, therefore, proposes a framework of system‐based malware detection through robust adversarial training that retains the identification with precise labeling. On adversarial samples, the suggested model applies projected gradient descent (PGD) and deep fool techniques when warping deep features with 11Multi‐Scale Dense Attention 1D dilated DCNN. The model employed in the accurate detection of malware classes is the optimal weighted federated deep learning (OWFDL). Local models and global models (gated recurrent neural networks (GRNN) with deep belief networks (DBN)) are used in this optimized weighted ensemble‐deep neural network (OWE‐DNN). In the proposed architecture, federated aggregation (FA) is used to accomplish model aggregation. The optimal parameter of the model is selected by chaotic brown bear optimization (CBBO). The implementation of this approach will be a major step forward in securing the industrial networks with important defensive measures to counter the current era of cyberattacks that threaten the fundamental systems of infrastructure. Additionally, Explainable artificial intelligence (XAI) is such that the human experts comprehend and interpret model detection, thereby raising the transparency and trust in the system. Finally, experiment evaluation used on two different datasets demonstrated the effectiveness of the proposed model, achieving 0.9884 accuracy on the SOMLAP dataset and 0.9985 on the Windows Malware dataset.

Abstract The planned IceCube-Gen2 radio neutrino detector at the South Pole will enhance the detection of cosmic ultra-high-energy neutrinos. It is crucial to utilize the available time until construction to optimize the detector design. A fully differentiable pipeline, from signal generation to detector response, would allow for the application of gradient descent techniques to explore the parameter space of the detector. In our work, we focus on the aspect of signal generation, and propose a modularized deep learning architecture to generate radio signals from in-ice neutrino interactions conditioned on the shower energy and viewing angle. The model is capable of generating differentiable signals with amplitudes spanning multiple orders of magnitude, as well as consistently producing signals corresponding to the same underlying event for different viewing angles. The modularized approach ensures physical consistency of the samples and leads to advantageous computational properties when using the model as part of a bigger optimization pipeline.

Abstract This paper presents a comprehensive analytical investigation into the maximum output power for single-mechanism and hybrid energy harvesting systems under both displacement-unconstrained and displacement-constrained conditions. The electrically lossless single-mechanism harvester is considered an ideal model and treated as a theoretical benchmark for comparison. The electrical losses often encountered in practice are taken into account for a more rigorous and realistic assessment. Under unconstrained operation, two power optimization strategies, the impedance matching and gradient descent methods, are examined. The results reveal that both approaches yield identical optimal solutions for single-mechanism devices, whereas the gradient descent technique proves more suitable for complex hybrid systems, for which conventional impedance matching does not reach the global optimum. Explicit analytical solutions are derived for the maximum output power and associated optimal conditions for both unconstrained and motion-constrained scenarios. It is shown that, while the hybrid architecture may not be substantially beneficial under unconstrained regimes, it offers significant advantages over single-mechanism harvesters in displacement-limited operation under strong excitation. In general, the power improvement depends on the effective figures of merit for the piezoelectric and electromagnetic transducers and the input excitation strength. In particular, the output power from a hybrid energy harvesting system can be up to twice as high as that of a single-mechanism generator when the two figures of merit are comparable. These findings provide critical insights to guide the design and implementation of efficient energy harvesters, the identification and realization of optimal system parameters, and the effective deployment of hybrid approaches in challenging operational environments.

The global shift to digital terrestrial television broadcasting (DTTB) from the conventional analogue has significantly transformed television culture, necessitating comprehensive technical and infrastructural evaluations. This study addresses the limitations of existing path-loss models for accurately predicting path loss in digital terrestrial television broadcasting in the UHF bands, motivated by the need for reliable, location-specific models that account for seasonal, meteorological, and topographical variations in Abeokuta, Nigeria. The study focuses on path-loss prediction in the UHF band using Ogun State Television (OGTV), Abeokuta, Nigeria, as the transmission source. Eight receiving sites, spaced 2 kilometers apart, were selected along a 16.7 km transmission contour. Daily measurements of received signal strength (RSS) and weather conditions were collected over one year. Seasonal path-loss models PLwet for the wet season and PLdry. For the dry season, models were developed using multiple regression analysis and further optimized using least squares (LS) and gradient descent (GD) techniques, resulting in six refined models: PLwet, PLdry, PLwet−LS, PLdry−LS, PLwet−GD, and PLdry−GD. Model performance was evaluated using Mean Absolute Error, Root Mean Square Error, Coefficient of Correlation, and Coefficient of Multiple Determination. Results indicate that the Okumura model provided the closest approximation to measured RSS for all the receiving sites, while the Hata and COST-231 models were unsuitable. Among the developed models, PLwet (RMSE− 1.2633, MAE − 0.9968, MSE − 1.5959, R − 0.9935, R2 − 0.9871) and PLdry−LS(RMSE− 1.1884, MAE − 0.7692, MSE − 1.4124, R − 0.9942, R2 − 0.9883) were found to be the most suitable models for the wet and dry seasons, respectively. The major influence of location-based elevation and meteorological data on path-loss prediction over digital terrestrial television broadcasting communication lines in Ultra-High-Frequency bands was evident.

We examine the solution paths of three transfer learning estimators within the context of linear models. Our analysis identifies a general solution path that characterizes how transfer learning estimators interpolate between the target and source estimators. This involves a change of basis and entry-wise weighting functions, with existing transfer learning methods identified as special cases. The proposed framework reveals connections and equivalences among transfer learning approaches, offering valuable insights for designing future estimators with improved control over the learning process. It extends beyond traditional penalized regression and gradient descent techniques and holds the potential for generalization to nonlinear cases. Extensive simulations validate the theoretically derived solution paths, and their practical utility is demonstrated by the improvement in risk prediction for end-stage renal disease in Hispanic populations.

Controlling the spike timing of individual neurons is a fundamental challenge with significant implications for treating neurological disorders. While much research has focused on neural models in low-noise scenarios, real-world applications, such as implantable therapeutic devices, must operate in noisy environments and address the diverse firing patterns of neurons. Leveraging the Izhikevich model, which captures a broad range of firing behaviors, this study proposes a novel method for spike timing control using Neural Stochastic Differential Equations (Neural SDE). The approach iteratively trains external currents to minimize both firing mismatches and timing errors through stochastic gradient descent and back-propagation techniques. Simulations demonstrate that the method achieves precise spike control across various neuron types and noise levels, including regular spiking, bursting, and fast spiking patterns. The approach remains effective even under strong noise perturbations, with particularly high precision observed in early spike events. Unlike conventional methods relying on deterministic dynamics or simplified models, the proposed Neural SDE framework directly accounts for biological noise and complex intrinsic dynamics. This enables the generation of neuron-specific control signals that align spike timings while adapting to individual firing characteristics. These results highlight the method’s generalizability and suggest its suitability for real-world neural control applications, including neuroprosthetics, adaptive stimulation, and closed-loop therapeutic systems.

This paper investigates the performance decline in complex systems, such as Unmanned Aerial Vehicles (UAVs), caused by unanticipated faults and external perturbations. To improve system resilience and achieve swift recovery without depending on fault detection, a passive Fault-Tolerant Control (FTC) approach is developed, combining Sliding Mode Control (SMC) with Radial Basis Function (RBF) neural networks. The RBF network, utilizing its robust approximation abilities, is applied to dynamically estimate system uncertainties, thereby alleviating the chattering issue typical of traditional SMC and minimizing its negative effects on system reliability and operation. Notably, this work addresses the challenges of instability and slow convergence often encountered in conventional gradient descent techniques for adjusting RBF network parameters. Instead, an enhanced Particle Swarm Optimization (PSO) method, incorporating an adaptive mutation mechanism (MPSO), is employed to effectively fine-tune the RBF network's critical parameters (centers and widths), resulting in improved convergence rates, learning performance, and parameter stability. The stability of the closed-loop system is thoroughly established using Lyapunov theory, ensuring that all signals remain bounded. Lastly, extensive simulations on a quadrotor UAV model under diverse fault conditions and disturbances are conducted to confirm the efficacy and highlight the advantages of the proposed MPSO-RBF-based adaptive sliding mode FTC approach over both conventional and standard adaptive SMC benchmarks.

This study addresses the development of methods and algorithms for evaluating delays based on machine learning models. Specifically, intelligent models and algorithms have been developed to assess sales delays in the economic domain. The process began with the preparation of a dataset for analyzing sales delays. Algorithms based on machine learning models such as Logistic Regression and Random Forest were developed to train on the prepared data. Furthermore, these algorithms were optimized using Gradient Descent techniques. Experimental results were obtained using the developed algorithm, and the model's accuracy was thoroughly analyzed. The results demonstrate that the Random Forest-based algorithm outperformed Logistic Regression, achieving higher accuracy and reliability in identifying potential delays. The proposed approach highlights the importance of data preprocessing, feature engineering, and model evaluation metrics such as Accuracy, Recall, Precision, and F1 Score in ensuring the effectiveness of delay detection models.

In this paper, a feed-forward artificial neural network (FFNN) is proposed to analyze the behaviour characterized by nonlinear advection-diffusion-reaction (ADR) equations. This approach uses a trial function that satisfies the initial and boundary conditions and depends on a neural network constructed to approximate the solution of the problem. Since the trial function contains unknown parameters, the solution process must be minimized by using efficient optimization techniques to obtain these parameters. Therefore, in this paper, the gradient descent (GD) and particle swarm optimization (PSO) techniques are proposed to address the minimization issue. The results obtained by combining artificial neural network (ANN) method with the optimization techniques have been compared and the advantages and disadvantages of the problems have been discussed. The results revealed that the proposed ANN techniques have produced accurate and reliable solutions by comparing the exact and available literature. Furthermore, these techniques are economical in terms of computational memory.

In optical networks, assessing the quality of transmission (QoT) of future demands is critical for optimized network operation. Given the recent advances in network telemetry and Big Data processing, there has been a growing trend to leverage the telemetry of existing connections to enhance the QoT estimation for future demands. Currently, QoT estimation techniques are categorized into three classes: physics-informed analytical models, machine learning (ML) models, and hybrid models, combining the first two classes. Hybrid models have the potential to achieve accurate QoT estimation with reasonable explainability. In this work, we propose a hybrid QoT estimation technique based on gradient transfer between wavelengths using a simplified network digital twin. The proposed approach relies on the GN model to create a starting point NDT (greenfield), which is then updated with signal-to-noise ratio (SNR) telemetry from deployed transponders (brownfield) using gradient descent techniques. By assuming that neighboring channels have similar performance, the method enables learning from adjacent lightpaths. We use well-established ML optimizers to enhance the convergence rate and reduce steady-state estimation error. Compared with classical analytical models, simulation results indicate improved accuracy with reasonable explainability in network scenarios with imperfect amplifier parameters.

Aerial manipulation provides an effective alternative to human labor in high-risk outdoor situations. Complex and variable environments demand the system to respond quickly with minimal latency to external disturbances. To address this challenge, we propose a learning-based variable whole-body model predictive controller designed to improve the adaptability and agility of the system through robotic arm-assisted motion. Given the limited onboard computing power, this low-level whole-body model predictive controller enhances computational efficiency without sacrificing accuracy by linearizing the highly coupled dynamics model and updating the linearized parameters in real-time. By incorporating updates of the disturbance values predicted by the Gaussian process into the linear model, the whole-body controller can swiftly react to perturbations. Additionally, it can employ robotic arm motions to perform agile maneuvers and counter disturbances, rather than merely adjusting the quadrotor's rotational movements. To further enhance agility and robustness, we train a high-level policy search using episode-based policy search and gradient descent techniques. For specific tasks and scenarios, this policy search can train a deep neural network to identify optimal decision variables that account for various wind disturbances for the low-level controller. We have carried out disturbance rejection and flip experiments on the aerial manipulation system in the wind tunnel, which demonstrate that the controller can operate stably and effectively under strong disturbance.

ABSTRACT The internet and communication industries have grown at a very quick pace, which has caused a massive increase in the volume of data and network size. This surge has given rise to a multitude of new attacks, posing substantial challenges for network security in effectively identifying breaches. To counteract these threats, intrusion detection systems (IDS) have been created, utilizing technology to scrutinize, monitor, and analyze network traffic and ensure the conservation of availability, confidentiality, and integrity. In networks with imbalanced traffic, malicious cyber‐attacks can easily go unnoticed within large volumes of regular data. This proficiency in concealing their presence poses a formidable obstacle for Network IDS in accurately and promptly detecting such threats. Despite extensive research efforts, conventional IDS proposed models are faced with persistent issues of enhancing detection accuracy and lowering false alarm rates, identifying emerging rare and zero‐day intrusion types. Previous research has also emphasized the problem of uneven distribution in network traffic, potentially leading to the misclassification of attacks. As a solution to these problems, this study proposed a multi‐model architecture that leverages attention mechanism‐based long short‐term memory (AM‐LSTM) and class‐wise focal loss‐based variational autoencoder (CWFL‐VAE), which both aimed to detect various forms of attacks, including rare or zero‐day attacks, while reducing false alarm rates and computational complexity. CWFL‐VAE was employed to handle imbalanced network traffic, focusing on minority classes to address the issue of misclassification; AM‐LSTM was used for classification, while the Adam gradient descent technique was employed to optimize the model. The proposed system performance was assessed using two datasets: NSL‐KDD, a benchmark dataset with skewed network traffic distribution, and CSE‐CIC‐IDS2018, featuring network traffic that is approximately 83% benign cases. CSE‐CIC‐IDS2018 was employed in assessing the performance of the model due to its recent release and incorporation of contemporary attack types, while the NSL‐KDD data functioned as a trustworthy benchmark, testing the model's implementation against findings in the literature. The research showed good performance with a low false positive rate of 0.12%, 99.37% accuracy, and 99.23% detection rate for the NSL‐KDD data. Similarly, the technique's detection rate, accuracy, and false positive rate for the CSE‐CIC‐IDS2018 data were 94.2%, 0.22%, and 92.39%, respectively. According to these findings, the recommended model was found to be competitive in terms of precision, rate of detection, and incidence of false positives when evaluated with existing methods.

In the world of big data, training large-scale machine learning problems has gained considerable attention. Numerous innovative optimization strategies have been presented in recent years to accelerate the large-scale training process. However, the possibility of further accelerating the training process of various optimization algorithms remains an unresolved subject. To begin addressing this difficult problem, we exploit the researched findings that when training data are independent and identically distributed, the learning problem on a smaller dataset is not significantly different from the original one. Upon that, we propose a stagewise training technique that grows the size of the training set exponentially while solving nonsmooth subproblem. We demonstrate that our stagewise training via exponentially growing the size of the training sets (STEGSs) are compatible with a large number of proximal gradient descent and gradient hard thresholding (GHT) techniques. Interestingly, we demonstrate that STEGS can greatly reduce overall complexity while maintaining statistical accuracy or even surpassing the intrinsic error introduced by GHT approaches. In addition, we analyze the effect of the training data growth rate on the overall complexity. The practical results of applying l2,1 -and l0 -norms to a variety of large-scale real-world datasets not only corroborate our theories but also demonstrate the benefits of our STEGS framework.

Summary Effective decision-making plays a crucial role in maximizing economic and global returns from an oil reservoir. When architecting strategies for exploration, development, and production throughout a reservoir’s life cycle, evaluating production forecasts through numerical simulations is essential. But the significant computational expense associated with simulating these strategies using reservoir models imposes time constraints and limitations. To address these challenges, researchers have explored methods to replace simulation with surrogate models (also known as proxies) that act as faster objective function estimators (FOFE), which can save time at each stage of the decision-making processes. We developed three analytical surrogates to handle multiple inputs with robustness to enhance an optimization algorithm. One surrogate used a genetic algorithm (GA) to parameterize inputs before calculating linear coefficients using the gradient descent technique. These surrogates underwent testing across 1,080 unique training conditions to assess their effectiveness in mitigating the simulation cost of a production strategy optimization. A method for comparing population samples provided by surrogates vs. those derived from optimization methods was proposed, and the relationship of these metrics with the surrogate’s prediction-to-true outcome correlation was investigated. Additionally, two methods for obtaining samples with surrogates were proposed, each with decision criteria of using surrogate with the Pearson’s correlation (r). An optimization algorithm that incorporates these surrogates was developed and applied to optimize well management in a real-case carbonated reservoir in the Brazilian presalt region. To evaluate the surrogate accuracy for different sample sizes during an optimization, three optimization population sizes were tested: 50, 30, and 20. The surrogates enabled optimization speedup of 48, 69, and 79% for the population sizes, respectively. When used to improve the net present value (NPV), the surrogate-aided optimization algorithms helped improve the base strategy’s NPV from USD 3.9 up to USD 4.3 billion compared with the benchmark.

Deep Equilibrium Unfolding Learning for Noise Estimation and Removal in Optical Molecular Imaging.

ABSTRACT Synthetic aperture radar (SAR) is widely used for ground observations owing to its all-weather and all-time capabilities. Sparse SAR reconstruction imaging, which can enhance SAR imaging performance, has emerged as a significant research area. The traditional sparse SAR reconstruction imaging method was successfully applied to raw SAR data processing. However, the computational complexity makes sparse reconstruction imaging difficult for large scenes. Meanwhile, because raw data are usually kept confidential, only limited data are available. To address the issue of how to use publicly available complex image data for sparse reconstruction and effectively improve image quality, this study proposes a joint L1- and F-norm edge-preservation-based sparse SAR reconstruction algorithm for complex images. The algorithm uses the gradient descent (GD) technique to optimize the L1-norm regularization and reconstruct the overall image information; the iterative soft thresholding algorithm (ISTA) to optimize the F-norm regularization and reconstruct the high-frequency image information; and fuses the two sets of reconstructed information to obtain SAR reconstructed images with characteristics such as edge preservation, clear texture, and high quality. The results show that, compared with traditional matched filtering (MF) imaging results, the joint L1- and F-norm results obtained for complex images have lower sidelobes, higher signal-to-noise ratio, and better target resolution ability; compared with the F-norm-based sparse SAR reconstruction imaging results, the joint L1- and F-norm method provides images with better edge texture clarity.

<div>In this work, the large-angle rotational movement and vibration suppression of a flexible spacecraft are carried out based on an adjustable system. First the spacecraft model is transformed into a canonical affine control form, then two fuzzy systems are used: The first (of Takagi–Sugeno type) estimates the feedback linearization control law as a whole, while the second (of Mamdani type) adjusts and stabilizes the control parameters using the gradient descent technique and based on the minimization of the control error rather than the tracking error. Stability results are presented in terms of Lyapunov’s theory, and simulation tests illustrate the significant transient robustness of the closed-loop system against perturbations, the accurate trajectory control, and vibration suppression of the flexible spacecraft. Consequently, as will be shown later, the error will stay confined and converges quickly to zero, confirming the smoothing property of the proposed method using fuzzy logic systems.</div>

This paper studies the hierarchical optimal synchronization(HOS) problem of heterogeneous nonlinear complex dynamic networks with input saturation constraints under the sequential decision-making mechanism. Firstly, a Stackelberg-Nash game model to solve the HOS problem is established, and the value functions of the major node and the minor nodes are introduced respectively. Then, the coupled HJB equations are derived based on the controlled network models. Furthermore, the expressions of optimal controllers for the major node and the minor nodes are theoretically given. Integral reinforcement learning algorithm based on value iteration method is designed, and the convergence, optimality, admissibility of iterative policy of the algorithm are analyzed. To implement this algorithm, the single-critic neural networks(NNs) are used to estimate the value functions, the gradient descent technique is employed to train NNs to toward the ideal weights. Finally, numerical simulations are provided to check the validity of the main results.

Abstract With a view to long-term qubit device operation, we report on a method to stabilize a silicon double quantum dot against slow drift in a two-dimensional gate-voltage space based on a current gradient-based feedback technique. We demonstrate that, unlike conventional single-axis feedback schemes, our method can maintain the double-dot potential configuration. We measure a feedback bandwidth of up to 300 mHz, consistent with the sampling rate and the digital filter cutoff frequency used in the experiment.

Convolutional neural networks (CNN) are widely used by researchers due to their extensive advantages over various applications. However, images are highly susceptible to malicious attacks using perturbations that are unrecognized even under human intervention. This causes significant security perils and challenges to CNN-related applications. In this article, an efficient adversarial training model against malevolent attacks is demonstrated. This model is highly robust to black-box malicious examples, it is processed with different malicious samples. Initially, malicious training models like fast gradient descent (FGS), recursive-FGSM (I-FGS), Deep-Fool, and Carlini and Wagner (CW) techniques are utilized that generate adversarial input by means of the CNN acknowledged to the attacker. In the experimentation process, the MNIST dataset comprising 60K and 10K training and testing grey-scale images are utilized. In the experimental section, the adversarial training model reduces the attack accuracy rate (ASR) by an average of 29.2% for different malicious inputs, when preserving the accuracy of 98.9% concerning actual images in the MNIST database. The simulation outcomes show the preeminence of the model against adversarial attacks.