Artificial neural networks are currently applied in a wide variety of fields, and they are near to achieving performance similar to humans in many tasks. Nevertheless, they are vulnerable to adversarial attacks in the form of a small intentionally designed perturbation, which could lead to misclassifications, making these models unusable, especially in applications where security is critical. The best defense against these attacks, so far, is adversarial training (AT), which improves the model’s robustness by augmenting the training data with adversarial examples. In this work, we show that the performance of AT can be further improved by employing the neighborhood of each adversarial example in the latent space to make additional targeted augmentations to the training data. More specifically, we propose a robust selective data augmentation (RSDA) approach to enhance the performance of AT. RSDA complements AT by inspecting the quality of the data from a robustness perspective and performing data transformation operations on specific neighboring samples of each adversarial sample in the latent space. We evaluate RSDA on MNIST and CIFAR-10 datasets with multiple adversarial attacks. Our experiments show that RSDA gives significantly better results than just AT on both adversarial and clean samples.

This study addresses the dynamic object detection problem for Unmanned Surface Vehicles (USVs) in marine environments, which is complicated by boat tilting and camera illumination sensitivity. A novel pipeline named “Seal” is proposed to enhance detection accuracy and reliability. The approach begins with an innovative preprocessing stage that integrates data from the Inertial Measurement Unit (IMU) with LiDAR sensors to correct tilt-induced distortions in LiDAR point cloud data and reduce ripple effects around objects. The adjusted data are grouped using clustering algorithms and bounding boxes for precise object localization. Additionally, a specialized Kalman filter tailored for maritime environments mitigates object discontinuities between successive frames and addresses data sparsity caused by boat tilting. The methodology was evaluated using the VRX simulator, with experiments conducted on the Volga River using real USVs. The preprocessing effectiveness was assessed using the Root Mean Square Error (RMSE) and tracking accuracy was evaluated through detection rate metrics. The results demonstrate a 25% to 30% improvement in detection accuracy and show that the pipeline can aid industry even with sparse object representation across different frames. This study highlights the potential of integrating sensor fusion with specialized tracking for accurate dynamic object detection in maritime settings, establishing a new benchmark for USV navigation systems’ accuracy and reliability.

Deep neural networks (DNNs) have gained prominence in various applications, but remain vulnerable to adversarial attacks that manipulate data to mislead a DNN. This paper aims to challenge the efficacy and transferability of two contemporary defense mechanisms against adversarial attacks: (a) robust training and (b) adversarial training. The former suggests that training a DNN on a data set consisting solely of robust features should produce a model resistant to adversarial attacks. The latter creates an adversarially trained model that learns to minimise an expected training loss over a distribution of bounded adversarial perturbations. We reveal a significant lack in the transferability of these defense mechanisms and provide insight into the potential dangers posed by $L_{\infty }$-norm attacks previously underestimated by the research community. Such conclusions are based on extensive experiments involving (1) different model architectures, (2) the use of canonical correlation analysis, (3) visual and quantitative analysis of the neural network’s latent representations, (4) an analysis of networks’ decision boundaries and (5) the use of equivalence of $L_2$ and $L_{\infty }$ perturbation norm theories.

Introduction. The prediction module generates possible future trajectories of dynamic objects that enables a self-driving vehicle to move safely on public roads. However, all modern prediction methods evaluate their performance only under urban conditions and do not consider their applicability to the domain of rural roads. This work examined the adaptability of existing methods to work under rural unstructured conditions and suggested a new, improved approach.Materials and Methods. As a solution, we propose to use a neural network that includes the following submodules: a graph-based scene encoder, a multimodal trajectory decoder, and a trajectory filtering module. Another proposed feature is to use an adapted loss function that penalizes the network for generating trajectories that go beyond the drivable area. These elements use standard practices for solving the prediction problem and adapting it to the domain of rural roads.Results. The presented analysis described the basic features of the prediction module in the rural road domain, showed a comparison of popular models, and discussed its applicability to new conditions. The paper describes the new approach that is more adaptive to the considered domain of study. A simulation of the new domain was performed by modifying existing public datasets.Discussion and Conclusion. Comparison to other popular methods has shown that the proposed approach provides more accurate results. The disadvantages of the proposed approach were also identified and possible solutions were described.

With the increasing volume of data transmitted over the network and the increasing number of devices connected to the network, there is an escalating need to facilitate network management. In this context, SDN (Software Defined Networks) technology was created to solve many network problems with its ability to separate the control plane from the data plane providing more flexibility and programmability to the network. Nevertheless, the centralization of network management in the SDN controller raises a new problem, in that the SDN controller is a bottleneck and its performance is critical to the whole network. In this paper we present using the DISC (Discrete Mathematics Instruction Set Computer) processor, a special processor developed in Bauman University, to improve the scalability and stability of SDN networks and enhance the performance of the SDN controller, seeking to give the controller the ability to work on bigger networks with more switches faster and with less power consumption.

Segmenting individual trees from airborne LiDAR point cloud data is critical for forest management, urban planning, and ecological monitoring but remains challenging due to complex natural environments, diverse tree architectures, and dense canopies. Traditional supervised methods rely on extensive, manually annotated datasets that are often impractical to obtain. In this study, we propose a novel self-supervised learning framework that eliminates the need for manual labeling by integrating transformation-invariant feature extraction, an energy-based segmentation loss, and soft clustering. The framework operates in two stages: a pretext task applies geometric transformations—rotation (from –45° to +45°), translation (between –1 and 1 units), and scaling (between 0.5 and 2.0)—to learn robust features, while an unsupervised segmentation step leverages an energy function that combines height, density, and slope attributes to cluster points corresponding to individual trees. We evaluated our approach on a high-density LiDAR dataset acquired from the Estonian Land Board (LAS format, version 1.4) comprising over 850,000 points. Our method achieves up to 87% convexity, 78% solidity, and an elliptical fit error as low as 0.12, substantially reducing over-segmentation compared to baseline clustering techniques. These results demonstrate that our self-supervised framework offers a scalable, label-free solution for precise tree segmentation, with significant advantages in accuracy and efficiency over traditional supervised methods.