This paper investigates beamforming design for Simultaneous transmission and reflection reconfigurable intelligent surface (Star RIS) assisted secure wireless communication with three operating protocols (energy splitting (ES), mode switching (MS), and time switching (TS)) with the assumption of the availability of imperfect CSI of two eavesdroppers at BS. We maximize the sum rate of legitimate users by designing both active and passive beamforming at BS and RIS, respectively, under the constraint of the limited maximum tolerable rate of eavesdroppers and a limited transmission power budget. The non-convex problem is solved by alternating optimization (AO) for active and passive beamforming optimization, and also successive convex approximation (SCA), and the penalty concave-convex procedure (PCCP). According to the simulation results, Star-RIS outperforms conventional RIS, ES mode outperforms MS and TS, and the performance is higher with lower channel estimation error.
Currency exchange rate forecasting has always been one of the important issues for economic activists. In this context, the stationary, and non-linear behavior of this variable and random walk claim mentioned in some empirical studies have made forecasting as one of the challenges, and concerns in the field of economics. The present study briefly classifies various currency exchange rates forecasting models and methods, then focuses on five deep learning methods, including Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and Reinforcement Learning (RL). For this purpose, the report of various studies on forecasting currency exchange rates using the above methods, with objectives, such as identifying the researchers in this field, the scope of studies, the scientific centers conducting studies, along with their geographical distribution, mixed methods used, studied currency pairs, forecasting periods, data frequency, evaluation criteria, obtained accuracy and features used for forecasting were studied. The study results will help future research in this field more effectively identify the research gaps with classified access to the previous studies, and define the topic and scope of future research to complete previous studies.
The popularity of cryptocurrencies has intensified the need for accurate volatility prediction models. This research proposes a novel approach to enhance conditional variance predictions for cryptocurrencies. By leveraging a feature selection technique that selects strong features based on price, return, and volatility cross-correlation analysis, we effectively select the most relevant features for an LSTM-based prediction model. To obtain initial volatility estimates, various GARCH family models (GARCH, EGARCH, and GJR-GARCH) were fitted to the dataset, with the best-fitting model selected based on minimum MSE and RMSE. Subsequently, the proposed MGF-LSTM model was applied to the top eight cryptocurrencies by market capitalization. Experimental results demonstrate that our model significantly reduces prediction errors, providing valuable insights for risk management and investment decision-making in the cryptocurrency market.
Visual Question Answering (VQA) is a complex task that requires models to jointly analyze visual and textual inputs to generate accurate answers. Reasoning and inference are critical for addressing questions that involve relationships, spatial arrangements, and contextual details within an image. In this study, we propose a model based on the BLIP framework, as a generative model, that enhances contextual understanding by incorporating dense-captions -detailed textual descriptions generated for specific regions within an image- along with spatial information extracted from the image. The model focuses on emphasizing visual information and extracting additional context to improve answer accuracy. Experimental results on the GQA dataset demonstrate that the proposed approach achieves competitive performance compared to state-of-the-art methods
In recent years, Convolutional Neural Networks (CNN) have been extensively used in machine learning algorithms related to images due to their exceptional accuracy. The multiplication-accumulation (MAC) in convolutional layers makes them computationally expensive, and these layers account for 90% of the total computation. Several researchers have taken advantage of pruning the weights and activations to overcome high computation bandwidth. These techniques are divided into two categories: 1) unstructured pruning of the weights can achieve heavy pruning, but in the process, it unbalances data access and computation processes. Consequently, compression coding for indexing non-zero data increases, which causes much more memory volume. 2) Structured pruning by the specified pattern prunes the weights and regularizes both computations and memory access but does not support high pruning amounts compared to unstructured pruning. In this paper, we proposed Quasi Structured Pruning (QSP) that profits from the high pruning ratio of unstructured pruning. The load balancing property in structured pruning has also been included in the QSP scheme. Implementation results of our accelerator using VGG16 on a Xilinx XC7Z100 indicate 616.94 GOP/s and 1437.7 GOP/s at just 7.8 watts power consumption for dense and sparse mode, respectively. Experimental results show that the accelerator is 1.38×, 1.1×, 2.77×, 2.87×, 1.91×, and 1.18× better in terms of DSP efficiency than previous accelerators in dense mode. As well, our accelerator has achieved 1.9×, 2.92×, 1.67×, and 1.11× higher DSP efficiency besides 4.52×, 5.31×, 10.38×, and 1.1× better energy efficiency than other state-of-the-art sparse accelerators.
The proliferation of fake news on social networks poses significant challenges for trust, security, and societal well-being. In this paper, we present a comprehensive study of fake news detection approaches and techniques, introducing a novel framework for news construction comprising four elements: news content, news context, news propagation, and news environment. We propose a new taxonomy of fake news detection techniques categorized into two primary types—individual methods (content-based, context-based, and propagation-based) and frameworks (hybrid and perception-aware methods). We highlight their strengths, weaknesses, and applicability by analyzing 14 state-of-the-art detection methods across platforms such as Twitter, Facebook, and Sina-Weibo. Furthermore, we address critical research gaps by identifying future directions, including early fake news detection, unsupervised learning, multimodal datasets, adversarial attacks on algorithms, multi-lingual platforms, and AI-generated content detection. Our findings and recommendations aim to serve as a foundation for developing new robust, scalable, and impactful fake news detection systems.
The rapid adoption of artificial intelligence (AI) technologies across diverse sectors has exposed vulnerabilities, particularly to adversarial attacks designed to deceive AI models by manipulating input data. This paper comprehensively reviews adversarial attacks, categorising them into training-phase and testing-phase types, with testing-phase attacks further divided into white-box and black-box categories. We explore defence mechanisms such as data modification, model enhancement, and auxiliary tools, focusing on the critical need for robust AI security in healthcare and autonomous systems sectors. Additionally, the paper highlights the role of AI in cybersecurity, offering a taxonomy for AI applications in threat detection, vulnerability assessment, and incident response. By analysing current defence strategies and outlining potential research directions, this paper aims to enhance the resilience of AI systems against adversarial threats, thereby strengthening AI's deployment in sensitive applications.
Wireless sensor networks (WSNs) are advanced tools for monitoring and controlling the environment, which are powered by a limited capacity battery, and the depletion of the sensor battery leads to the end of the life of the network life. Therefore, it is crucial to use protocols that are energy efficient. In this paper, using a new cross layer model based on distance from base station (BS) and (Time Division multiple access) TDMA, the optimal use of available resources and increasing of the life time of the network are discussed. By modifying the method of selecting the cluster headers (CHs), the selection of low energy CHs, which are far from the BS, has been prevented. It also balances the transmission of data packets in different clusters, resulting in fair energy consumption between sensors. By comparison with the ATEER model in [12], the proposed model has reduced energy consumption by 45%, increased the life time of the network by 67% and increased packets sent to the BS by 15%.
Nowadays, with the impressive development of cellular networks and other wireless communication technologies, the public concerns about possible health effects of exposure to electromagnetic radiations are increased. In this paper, we investigate some radio frequency (RF) exposure levels measured in the urban areas. These measurements include all RF radiations of various electromagnetic wave sources including those from broadcasting stations to cellular base stations. The exposure levels are incident power density values measured during 18 months at 930 locations distributed over 7 major cities of Iran. These results are compared with the relevant guidelines of non-ionizing radiation given in the Iran national standard no. 8567, to confirm that the measured radiation exposure levels are satisfactorily below the defined limits.
With the expansion of artificial intelligence-based products, including IOT systems, as well as rap id adoption of algorithms in business and global society, there is a growing concern to preserve public interest, and the quality of these systems for practical use is of high importance. Smart AI wearable systems are among these products; these devices have various sensors such as accelerometer, gyroscope and other sensors that identify and categorize a person's physical activities, receive environmental and physiological information and process them using artificial intelligence, presenting useful information including physiological function, health and human behavior. As a result, this field can be valuable for scientific research, medicine, sports, industries and even daily life. For widespread use of smart AI wearable products, it is necessary to guarantee the quality of these products, and for this purpose, regulatory frameworks and evaluation criteria have been established in these systems. In this article, we used WSDM data to classify the activities performed by a user using CNN2 network and subsequently evaluated and tested Smart AI wearable systems in order to standardize these products. These tests include bias tests, black tests, robustness and generalization. The results of these tests have been discussed and evaluated at the end of article, which indicate successful design and implementation of this product. With the expansion of artificial intelligence-based products, including IOT systems, as well as rap id adoption of algorithms in business and global society, there is a growing concern to preserve public interest, and the quality of these systems for practical use is of high importance. Smart AI wearable systems are among these products; these devices have various sensors such as accelerometer, gyroscope and other sensors that identify and categorize a person's physical activities, receive environmental and physiological information and process them using artificial intelligence, presenting useful information including physiological function, health and human behavior. As a result, this field can be valuable for scientific research, medicine, sports, industries and even daily life. For widespread use of smart AI wearable products, it is necessary to guarantee the quality of these products, and for this purpose, regulatory frameworks and evaluation criteria have been established in these systems. In this article, we used WSDM data to classify the activities performed by a user using CNN2 network and subsequently evaluated and tested Smart AI wearable systems in order to standardize these products. These tests include bias tests, black tests, robustness and generalization. The results of these tests have been discussed and evaluated at the end of article, which indicate successful design and implementation of this product.
—Currently, the development of the coronavirus as a pandemic and its global spread are a major concern for our society and the international community. In recent years, however, a growing number of people have transferred their primary source of news and information to social networks. So, the broad transmission of inaccurate and misleading information on social media is significant for the majority of politicians. Not only are we fighting against COVID-19, but also a "infodemic." To address this, on COVID-19, we have collected and released a labeled dataset of 7,000 Persian social media postings of true and fake news. Several languages, including Arabic, English, Chinese, and Hindi, have recognized Covid 19 fake news. This study utilizes a deep neural network approach to simplify feature extraction, develop a strong ability to learn, and automatically discover features compared to typical machine learning approaches, as well as a novel approach to improving outcome using a deep neural network. The genetic algorithm and reinforcement learning are provided for setting and optimizing the hyper-parameters of the deep learning algorithm, which has led to better outcomes than previous research and achieved an accuracy rate of 0.92 percent.
This research presents a novel portfolio optimization framework using deep reinforcement learning (DRL). Traditional methods rely on static models or single-agent strategies, which struggle with market dynamics. We propose a dynamic system to address this by selecting the best-performing DRL agent based on recent market conditions. The framework evaluates five DRL agents, A2C, SAC, TD3, DDPG, and PPO, allocating portfolio weights based on short- term performance. A selection mechanism identifies the top agent using cumulative returns over the prior ten days, leveraging multiple agents' strengths. This adaptive approach embraces the philosophy that no single strategy consistently outperforms in all market conditions, making flexibility and continuous learning essential for robust portfolio management. Backtesting on Dow Jones data shows our method enhances cumulative returns and risk- adjusted performance, achieving an 11.43% average annual return, 38.29% cumulative returns, and a 0.832 Sharpe ratio, outperforming individual DRL agents.
Text summarization is the process of condensing a source text while retaining its key points, tailored to a specific audience or task. The research extractive summarization, where each news article was segmented into individual sentences. Each sentence underwent processing through the ParsBERT algorithm. Subsequently, an attention layer combined the sentence weights with the Bidirectional GRU algorithm's output to extract summarized sentences for labeling. The dataset comprised over 175,000 articles sourced from reputable Persian news agencies (ISNA-TASNIM), covering various topics such as science, politics, and sports. Evaluation of the summarization techniques was conducted using Rouge metrics. The results of the investigation revealed precision values of 0.7923 (Rouge-1), 0.7613 (Rouge-2), and 0.8582 (Rouge-L). The study also evaluated the effectiveness of Gated Recurrent Unit (GRU) algorithms in extractive summarization by integrating its architecture with the attention network. The results demonstrated an improvement in news text summarization compared to other deep learning hybrid algorithms.
According to today's statistics, more than half a billion vehicles are moving in the world and inspection and monitoring is one of the basic needs of any traffic system. All cars have an identification number or the same license plate as their primary ID, which today is one of the most suitable vehicle authentication tools. In this paper, the high capacity of deep neural networks in learning license plate identifiers is used. The proposed model of this paper has two stages of highlighting the license plate and reading the ID. In this regard, for highlighting, the combination of YOLO and XGBOOST network is used in encoder-coder network. The proposed model is evaluated on the FZU Cars dataset and based on the results of the experiments, the proposed model has a higher accuracy than the basic methods.
In recent years, due to the high availability of documents through the Internet, plagiarism is becoming a serious issue in many fields of research. Moreover, the availability of machine translation systems facilitates the re-use of textual content across languages. So, the detection of plagiarism in cross-lingual cases is now of great importance especially when the source and target language are different. Various methods for automatic detection of text reuse have been developed whose objective is to help human experts investigate suspicious documents for plagiarism cases. For evaluating the performance of theses plagiarism detection systems and algorithms, we need to construct plagiarism detection corpora. In this paper, we propose an English-Persian plagiarism detection corpus comprised of different types of paraphrasing. The goal is to simulate what would be done by humans to conceal plagiarized passages after translating the text into the target language. The proposed corpus includes seven types of paraphrasing methods that cover (but not limited to) all of the obfuscation types in the previous works into one integrated CLPD corpus. To evaluate the corpus, an extrinsic evaluation approach has been applied by executing a wide variety of plagiarism detection algorithms as downstream tasks on the proposed corpus. The results show that the performance of the algorithms decreases by increasing the obfuscation complexity.
Production is a key component of every nation’s economy, yet the manufacturing sector faces major challenges and opportunities due to rapid digital transformation. Many companies have not fully adapted to these technological shifts, limiting their ability to gain competitive advantages. Research indicates that integrating digital approaches into production processes can enhance efficiency and create significant value, turning digital transformation from a strategic recommendation into a necessity. However, there is still limited guidance on how to systematically assess the digital maturity of manufacturing firms and support their progress toward higher maturity levels. This study aims to develop a comprehensive framework for evaluating digital maturity in the manufacturing sector. Drawing on both literature and empirical data, the framework was designed and validated to help organizations understand their current digital status and identify areas for improvement. It defines evaluation domains, maturity levels, and assessment criteria, along with a structured evaluation method to guide practitioners in achieving higher levels of digital transformation.
A maturity model, is a tool that can be applied to assess the “As-is” situation regarding specific dimensions. Inherent to the growing business-related data creation and storage in many enterprises, data analytics plays a challenging key role in business success. The more an organization is data dependent, the more the data analysis maturity model is crucial. In this paper, we review data analysis maturity models and propose an easy-to-use data analysis maturity model with tools to assess maturity level. As a case study, the proposed model has been successfully applied to a financial/banking enterprise, in which data analysis is crucial and can gain deeper insights into business operations, improve decision-making processes, manage risks more effectively, and ultimately drive better financial performance. The proposed model is beneficial to be used for any enterprise, especially data-driven ones whose data analysis enhances their competency.
One of the significant problems of handicraft makers is the inability to trade with customers directly. They are forced to sell their crafts to intermediaries, which increases the final price. To avoid such an issue, peer-to-peer marketplaces are built to connect producers and customers directly. The ownership and control of these marketplaces are still in the hands of specific parties. Decentralized markets overcome these drawbacks at a reasonable level. Blockchain has become a great tool to develop decentralized markets due to solving the shortcomings of decentralized systems. We have studied, categorized, and discussed existing decentralized e-commerce platforms. To cover the shortcomings of decentralized e-commerce platforms, we designed a conceptual marketplace for trading handicrafts along with an insurance mechanism for deliveries, seller's reputation, and a guarantee system for producer's verification. Comparing the handicraft marketplace with other similar decentralized markets proves offering better and more service to marketplace actors is achievable.
Resource allocation and task scheduling in edge computing environments are crucial for optimizing overall system performance. This paper introduces a multi-stage heuristic approach that combines local search, Genetic Algorithm (GA), and Particle Swarm Optimization (PSO) to address the complexities of resource management in a three-layer architecture consisting of IoT devices, edge nodes, and cloud servers. The proposed method aims to minimize latency, reduce energy consumption, and maintain load balance by intelligently distributing computational tasks among the available nodes. The multi-objective optimization framework dynamically adapts to changes in workload and network conditions, thereby enhancing the efficiency of the system. Experimental results show that the proposed approach outperforms traditional methods in terms of reducing response time and energy usage while achieving a balanced load distribution across edge nodes, making it an effective solution for real-time and resource-intensive applications in edge computing environments.
A reflectarray optimized through a Generative Adversarial Network (GAN) is demonstrated. This design focuses on the impact of the top layer on the reflection phase and utilizes the correlation between phase distribution and the direction of the reflected beam. Six programmable subcells are optimized to accommodate two incident angle waves simultaneously. Silicon substrate is exploited making the design compatible with integrated circuits. The far-field analysis indicates that for incident angles of 19.471° and 41.81°, as well as their vicinity, the reflectarray effectively redirects the incoming waves to reflect towards near-normal direction to its surface. This suggests a near independence of the deflection angle from the incident angle within a specific angular range, making the proposed reflectarray a planar THz beam collimator. The proposed subcells achieve a reflection phase range of 342°. The return losses for the incident angles of 19.471° and 41.81° are 1.9 dB and 1.4 dB, respectively. For a finite reflectarray measuring 15λ×5λ, the pattern gain and fractional bandwidth are reported as 19.44 dB and 24.8% for the incident angle of 19.471°, and 19.17 dB and 29.9% for the incident angle of 41.81°. This denotes an excellent wideband behavior for the proposed single-layer pixelated reflectarray.
