Belarusian State University of Informatics and Radioelectronics

Objectives. Development of methods for design compact and efficient neural networks for image recognition tasks, as well as their hardware implementation based on FPGA. Methods. The paper proposes the concept of a learnable two-dimensional separable transformation (LST) for designing feedforward neural networks for image recognition tasks. A feature of the LST is the sequential processing of image rows by a fully connected layer, after which the resulting representation is processed by columns using second fully connected layer. In the proposed architecture of a feedforward neural network, the LST is considered as a feature extractor. The hardware implementation of LST-based neural network is based on the concept of in-place computing (shared memory for storing source and intermediate data), as well as using a single set of computing cores to calculate all layers of the neural network. Results. A family of compact neural network architectures LST-1 is proposed, differing in the image embedding size. Experiments on the classification of MNIST handwritten digits have shown the high efficiency of these models: the LST-1-28 network achieves 98.37 % accuracy with 9.5 K parameters, and the more compact LST-1-8 shows 96.53 % accuracy with 1.1 K parameters. Testing of the LST-1-28 hardware implementation confirms the architecture's resistance to parameter quantization errors. Conclusion. The proposed concept of a learnable two-dimensional separable transformation provides the design of compact and efficient neural network architectures characterized by: a small number of learnable parameters, high recognition accuracy, and the regular structure of the algorithm, which makes it possible to obtain their effective implementations based on FPGAs.

Objectives. The purpose of the study is to examine the operational characteristics of a digital circuit designed to analyze the output signal frequency of a configurable ring oscillator within a fixed measurement window. Methods. Methods of digital device synthesis and analysis were employed, including implementation on programmable logic integrated circuits (FPGAs), fundamentals of digital circuit design, and methods for analyzing normally distributed random variables. Results. A digital circuit for recording the period of a configurable ring oscillator, depending on the measurement time window and the value of its configuration has been developed. Experimental studies of the generated signal periods were conducted using Xilinx ZYNQ 7000 series FPGAs. It was demonstrated that upon repeated period measurements the bits of the recording counter can be categorized into three groups: group G 2 : stable bits retaining constant values across all measurements; group G 1 : weakly stable bits exhibiting minor distortions; group G 0 : strongly unstable bits with a distortion probability approaching 1 between measurements. It was hypothesized that group G 0 represents the digitized noise component of the measured period value. Due to numerous independent factors (components within the configurable ring oscillator and digital recorder, supply voltage deviations, die and ambient temperature variations, quantization errors, etc.), it is assumed that this noise component follows a normal distribution. Analytical proof established that a normally distributed variable, quantized using multi-bit binary numbers under specific values of mathematical expectation μ and standard deviation σ , generates only two groups: G 2 and G 0 . It was proven that the probability of a '1' appearing in any bit of group G 0 approaches 0,5, and the group size can be estimated as 3+ ⌊log 2 σ⌋. The bits of group G 1 can be converted to group G 2 using various methods, such as maximum likelihood estimation or by normalizing each measurement value to a theoretically justified separation into G 2 and G 0 . The values of group G 2 bits can be interpreted as a response to a challenge defined by the ring oscillator circuit configuration and measurement window, forming a novel type of multi-bit Physical Unclonable Function (PUF) characterized by high stability. Conversely, the G 0 bits can serve as single-bit sources of random variables with near-uniform distribution, providing a foundation for building random number generators. Conclusion. The obtained results can be utilized in embedded systems for providing unclonable identification of digital devices and for random data generation. The application of a synchronous binary counter as a frequency recording circuit for a configurable ring oscillator opens new avenues for designing multibit physical unclonable function architectures with enhanced performance metrics in terms of stability, uniqueness, and randomness.

The current technological race to increase the volume of received, processed, and transmitted information plays a crucial role in the security of any country, as these areas are fundamental for the deployment of complex linguistic and experimental models, or frontier models, used in digital ecosystems and military affairs. This is particularly true for communications equipment and the strength of their cryptographic encryption. Compromising transmitted information, hidden from official subscribers of a closed radio network, can cause far greater damage than its failure. Given the rapid pace of change and innovation, countries without their own manufacturing capabilities are forced to manufacture digital encryption modules in other countries, which carries the risk of introducing hardware Trojans. This article describes the results of software testing of a neural network capable of detecting information compromise in an AES-256 (Advanced Encryption Standard) encryption module based on the analysis of received and transmitted information without a “golden reference”.

This paper examines the automation of medical image processing for diagnosing arterial hypertension using artificial intelligence and computer vision technologies. A software component has been developed that automatically extracts and structures information from visual representations of medical documents (including biochemical analysis results, complete blood counts, and 24-hour blood pressure monitoring data), minimizing errors and accelerating the process of entering and interpreting medical information. Algorithms for image preprocessing (increasing image resolution, noise removal, and tilt correction), segmentation, and text recognition were developed and tested using the Real-ESRGAN and EasyOCR neural network models. Particular attention was paid to improving text recognition quality in the presence of characteristic artifacts that arise when scanning or photographing documents. CER and WER metrics were used to evaluate quality, and the module's performance was assessed with and without superresolution. The results of the study confirmed the effectiveness of the proposed approach and demonstrated that the integration of Real-ESRGAN technology improves the accuracy of medical image processing in the presence of significant noise and low-resolution source data. The practical significance of the study lies in simplifying and accelerating the process of diagnosing hypertension and creating the basis for a personalized approach to patient treatment.

This article examines the impact of modern globalization processes on intellectual property risk assessment in the format of transnational interactions. Globalization creates new challenges related to the interaction of various legal systems, cultural norms, and economic factors, significantly complicating the process of identifying, assessing, and managing intellectual property risks. Emphasis is placed on the need to consider key aspects, including legal paradigms, cultural differences, and technological innovations that shape risks. Diagrams of interactions between participants in the intellectual property risk assessment process, both with and without the use of blockchain technologies, are developed. Risk management strategies are considered, emphasizing the importance of corporate social responsibility and the need for flexibility and adaptability in business processes. The importance of an integrated approach to risk management in a rapidly changing external environment is demonstrated, and practical recommendations are offered for organizations seeking to effectively protect their intellectual property rights in the face of increasing international competition.

Wide band gap (WBG) oxide and metal nanocomposites can possess bifunctionality from combining tightly coupled nanoobjects with different physicochemical properties. Adjusting synthesis conditions tunes these properties through modulating the process–morphology–function relationship. However, the controllable synthesis of such nanocomposites and their related applications are still underexplored. Here, we present a novel process flow to synthesize crystalline ZnO nanosheaves dotted with silver nanoparticles. The uniqueness of our strategy lies in the use of a silver mask for vertical growth of ZnO nanosheaves and thermal evaporating/dewetting Ag film to form a photocatalytic/plasmonic heterostructure. Upon combining a huge specific surface area and nanocrystallinity of ZnO nanosheaves, we enabled its surface-enhanced Raman scattering (SERS)-activity free of plasmonic components, yet their Ag modification resulted in improving detection limit in relation to Ellman’s reagent. Ag/ZnO nanosheaves showed dramatic photocatalytic activity to clean SERS-active surface. The systematic approach to synthesize Ag/ZnO heterostructure holds great promise in practical applications associated with interest in both photocatalytic and plasmonic properties.

Accurate trajectory prediction of multiple agents is a critical task in the fields of autonomous driving, human-computer interaction, and behavior analysis. However, the dynamic and interactive nature of agent behavior poses significant challenges, since it requires the formation of complex spatio-temporal dependencies and dynamically evolving interactions between agents. A novel approach is proposed for modeling dynamic relational graphs, the core component of which is the attention focus block, taking into account the relative positions of graph-based agents. By considering objects in a scene (e.g., vehicles and road elements) as graph nodes and their interactions as edges, the proposed approach effectively captures both local and global dependencies in a scene and makes a prediction about the future trajectory. The presented approach is evaluated using the Argoverse1 trajectory prediction dataset. Experimental results show that this model outperforms existing methods.

Features of formation of amplitude-frequency characteristics of resonant amplifier stages with double autotransformer inclusion are considered. The amplitude-frequency characteristic of the resonant cascade was simulated using the Multisim circuit design program. The transfer function of the circuit is analyzed and the factors influencing the formation of the frequency response of the cascade in a wide frequency range are revealed. Attention is drawn to the undesirable distortion of the shape of the amplitude-frequency response of the cascade outside the operating frequency band, leading to a decrease in attenuation and deterioration of selectivity on the side channels of reception. Recommendations are given for improving the shape of the amplitude-frequency characteristics outside the operating frequency band by parallel connection of autotransformers.