The deployment of deep learning models for face detection in edge devices presents challenges related to computational demands, resource constraints, and security. This paper will tackle the optimization for the distribution of face detection with different trained models to various hardware capabilities with the implementation of 4 secure communication protocols. To validate our proposal, we perform matrix evaluations across different NVIDIA edge devices and GPU-enabled PCs (GTX 1070Ti and RTX 2070), measuring performance metrics such as latency, throughput, and resource usage. The results demonstrate that our framework effectively balances accuracy, speed, and efficiency, providing optimal model-protocol combinations for edge devices.
The application of machine learning and deep learning to intrusion detection systems (IDSs) enhances their ability to detect and respond to sophisticated cyber threats efficiently and effectively, providing a robust defense mechanism in the ever-evolving landscape of cybersecurity. However, many environments where IDSs are deployed, such as IoT devices or edge computing nodes, have experienced several challenges with limited computational power, memory, and storage. In this paper, we leverage Explainable Artificial Intelligence (XAI) for feature selection and Knowledge Distillation (KD) to improve performance and decrease the complexity of the model for limited-resource devices. Specifically, XAI can identify features with little to no impact on the predictions, which can be removed to reduce the dimensionality of data and simplify the model. Besides, we leverage the KD technique to transfer the knowledge from a larger model to a small one to obtain a lightweight yet powerful IDS. Moreover, a new optimizer is also utilized to achieve high accuracy and robustness, while maintaining the speed and efficiency necessary for real-time intrusion detection. The experiments on three common IDS datasets, including CIC-IDS2017, UNSW-NB15, and NSL-KDD, prove our proposed method’s effectiveness with high accuracy.
As online services continue to grow, Distributed Denial of Service (DDoS) attacks pose significant threats to network systems. While Machine Learning (ML) techniques have increased adaptability in DDoS detection, they still face challenges in capturing complex patterns in high-dimensional or non-linear network data. Deep Learning (DL) models, such as Multilayer Perceptrons (MLP), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks, have advanced detection capabilities but encounter limitations when dealing with spatial and temporal data, sequential information, and computational efficiency. To address these challenges, this study proposes the use of the Transformer model for DDoS detection, enhancing its ability to identify intricate patterns in network traffic. Additionally, we introduce a novel packet-based data preprocessing method that preserves detailed packet-level information while maintaining a comprehensive view of data flows. Together, these contributions aim to significantly improve the accuracy and performance of DDoS detection systems.
In the digital age, electronic devices and digital products have become indispensable. They play an important role in both personal and professional settings for millions of people worldwide. As our dependence on technology increases, so does the urgency of ensuring cybersecurity. Cyber threats, particularly those attacking individual users and organizations, have become more frequent, complex, and difficult to detect. Among these, malicious JavaScript code is one of the most popular tools used by attackers to gain unauthorized access to systems and compromise sensitive information. Detecting these threats requires advanced and creative solutions. This paper introduces a graph learning approach, followed the rules of Markov chains, aims at detecting and mitigating malicious JavaScript attacks. Our approach takes the advantage of the structural properties of JavaScript code, transforming it into graph representation, which allows for a more effective identification of abnormal behavior. To evaluate the performance of our model, we use two popular datasets: Hynek Petrak’s dataset, which includes a variety of malicious JavaScript samples, and the SRILAB dataset, which contains benign JavaScript examples. Our method is compared with other leading graph learning models, and the results demonstrate a notable increase in detection accuracy. This approach offers a powerful tool to strengthen cyber security defenses in facing of evolving digital threats.
This paper investigates the secrecy performance of reconfigurable intelligent surfaces (RISs)-unmanned aerial vehicles (UAVs) within a 6G communication system. Specifically, a base station employing rate-splitting multiple access (RSMA) transmits both common and private messages to Internet of Things (IoT) devices (IDs) via a UAV relay (UR). Additionally, a mobile UAV eavesdropper (UE) is considered, capable of moving to intercept signals relayed by the UR. Therefore, we address the challenge of mobile UE position tracking to enhance physical layer security (PLS). We propose a framework that integrates real-time UE tracking using a particle filter (PF), a sequential Monte Carlo, for estimating the UE’s state. By continuously updating the UE location estimates, the considered system enables a proactive security measure, such as friendly jammers, to minimize the risk of intercepted communications. Simulation results demonstrate that accurate UE tracking significantly improves secrecy capacity and reduces the likelihood of successful eavesdropping.
This paper introduces a novel image captioning model designed to effectively combine the geometric and appearance features of objects with linguistic context. The model employs dual self-attention mechanisms to capture both spatial and visual relationships among objects, along with a cross-attention module to align these features with corresponding textual descriptions. The proposed has been evaluated on two Vietnamese image captioning datasets, showcasing its ability to generate accurate, contextually rich captions that outperform several state-of-the-art models. The proposed approach addresses the unique challenges of Vietnamese language structure in image captioning tasks, demonstrating robustness in handling complex interactions between visual and linguistic features. The results indicate that the model effectively bridges the gap between image content and textual representation, making it a promising solution for image captioning applications in Vietnamese. This model offers a significant contribution to the field, particularly for localized applications where understanding the meaning of language and context is essential to generate high-quality image descriptions that are both relevant and informative.
Anonymous communication in the sixth generation (6G) networks plays an essential role in enabling private and secure communication in applications such as digital twins, distributed computing, and the Internet of Things (IoT). As 6G networks continue to expand, conventional classical communication protocols face limitations in ensuring user anonymity and message confidentiality. Integrating advanced anonymous communication protocols is therefore crucial to meeting the stringent privacy and security demands. Quantum communication protocols can overcome these challenges and enhance secure information sharing based on the principle of quantum mechanics. In this paper, we propose an anonymous quantum group communication (AQGC) protocol for quantum linear network, designed to ensure the anonymity of both the sender and the recipients. This protocol allows the sender to transmit messages to multiple receivers simultaneously and anonymously. By utilizing entanglement swapping with Bell pairs, we achieve efficient, anonymous, and secure long-distance message transmission without requiring direct interaction between participants. We evaluate the protocol’s performance by analyzing its anonymity, correctness, and scalability. Our work shows the practicality of quantum communication protocols to address challenging privacy and security requirements, laying the groundwork for advanced communication systems in future quantum networks.
Abstractive Question Answering (AQA) is a pivotal task in Natural Language Processing (NLP) that requires AI models to not only comprehend the context and the question but also generate accurate, clear, and comprehensive answers. This task holds particular significance in the medical field, where AI-driven question answering has numerous applications, such as assisting with patient queries, supporting clinical decision-making, and improving access to medical information. In particular, pregnancy health is a critical area within health care AI, where there is increasing focus on leveraging AI to enhance outcomes in pregnancy-related care. However, medical question answering presents unique challenges due to the complexity of symptoms, which are often ambiguous or overlapping across different health conditions. Thus, the development of high-quality datasets is crucial for building and evaluating healthcare AI systems. In this paper, we introduce a closed-domain abstractive question answering dataset in Vietnamese, specifically focused on pregnancy health. This dataset is designed to address the distinct challenges of medical QA in a low-resource language and aims to support further research and innovation in healthcare AI, particularly in the context of pregnancy-related care.
Cardiovascular diseases (CVDs) are regarded as one of the leading causes of mortality worldwide, making the early detection of arrhythmias crucial for minimizing their impact. Electrocardiogram (ECG) signals are widely used to monitor heart health, and the automation of arrhythmia classification is deemed essential for enhancing diagnostic accuracy in clinical settings. In this study, a comprehensive framework for automated arrhythmia detection is presented, featuring an advanced preprocessing pipeline that combines band-pass filtering, notch filtering, and wavelet transform to eliminate noise from raw ECG signals effectively. Following noise removal, the R-peaks of the signals are detected, and based on these R-peaks, optimized window settings are applied for heartbeat segmentation, ensuring high-quality data for further analysis. The processed heartbeats undergo classification through several deep learning architectures: convolutional neural networks (CNN), bidirectional long short-term memory networks (BiLSTM), and a hybrid CNN+BiLSTM model. Model performance is optimized through hyperparameter tuning using grid search and 5-fold cross-validation. Ultimately, exceptional results are achieved by the proposed CNN+BiLSTM model, with an accuracy of 99.435
Through online social networks, members can create events and invite others to join. However, predicting the range of members who will attend upcoming events is a challenging problem. This paper studies the problem of event scale prediction in event-based social networks (EBSNs), which is crucial for event organizers to host events. Firstly, we investigate the characteristics of EBSN to extract three main factors of events. Then, we generate features based on three factors and use the K-means method to cluster participants of events into three different event scales. Next, we obtain a dataset; each sample of this dataset represents one event by a vector of 18 features and one scale label. We take the obtained dataset to train predictive classifiers. Finally, when a new event is coming, we use trained models to predict the scale of this event. Extensive experiments were conducted on a real crawled Meetup event dataset, and results have illustrated the effectiveness of our method in terms of accuracy. Moreover, we find that the content factor is the most critical factor that affects the performance of models and the decisions of members.
With the rapid convergence of artificial intelligence and mobile edge computing, edge intelligence is rapidly developing towards multi-dimensional perception and endogenous intelligence. As a promising enabling technology in edge intelligence, Artificial Intelligence Generated Content (AIGC) can provide mobile users with a variety of creative generation and personalized services, and further empower the paradigm shift of edge intelligence towards edge endogenous intelligence with its powerful reasoning and generalization capabilities. However, the large-scale parameters and fine-tuning process of AIGC make the existing task offloading schemes inappropriate in AIGC-empowered edge intelligence due to their high energy consumption and long waiting time. In this paper, we propose an energy-efficient task offloading scheme for AIGC-empowered edge intelligence. By considering the time cost and energy cost of the model fine-tuning process, the proposed scheme can accurately select the most suitable edge server and model to perform AIGC tasks, which further improves energy efficiency, service quality and user experience. Simulation results demonstrate that the proposed framework effectively minimizes the AIGC service failure rate and user waiting time compared to the baseline while ensuring efficient energy utilization.
This paper presents a novel approach to enhance simultaneous wireless information and power transfer (SWIPT) using an active intelligent reflecting surface (IRS), aimed at optimizing spectral efficiency under energy harvesting constraints. Specifically, an active IRS equipped with amplifying/reflecting elements is adopted to facilitate the transmitter’s simultaneous information and energy transmission to multiple power-splitting receivers, with the objective of maximizing the sum rate under power constraints. To address this, we develop a joint optimization framework using semidefinite relaxation (SDR) and successive convex approximation (SCA) that optimize the beamforming vectors and power-splitting ratios under the special case of fixed IRS amplification factors, phase shifts. Finally, simulation results demonstrate that the proposed solution algorithm converges and obtains the high sum rate results for the considered active IRS-aided SWIPT system.
In many speech applications, the essential core problem of signal processing is extracting the desired target speaker from the noisy mixture, which contains the original speech components and background noise, third - party talker, interferences and non- directional noise sources. Speech enhancement method, which utilizes the spatial information about geometry, the direction of arrival (DoA) of an interesting signal, the characteristics of the surrounding environment to achieve noise suppression and speech enhancement at the same time. The microphone array (MA) beamforming has attracted the attention of scholars, researchers and engineering to develop multi-channel signal processing algorithms, especially in acoustical instruments, hearing aids, cochlear implant, mobile phone, teleconference system, surveillance devices and smart home. Minimum Variance Distortionless Response (MVDR) beamformer is an efficient solution for extracting the desired target speaker at a specified direction while minimizing the total output noise power at the output signal of the beamformer. However, in real-life cases, MVDR beamformer’s performance is often corrupted, due to the displacement of MA distribution, the error of sampling frequency and the preferred DoA, the tolerance of MA amplitudes or the difference of microphone sensitivities. In this contribution, the author proposed applying an additive gain function to improve the overall MVDR’s evaluation in real-life environments. The demonstrated experiment confirmed the effectiveness of the author’s suggested technique in reducing the speech distortion to 5.0 dB, enhancing the speech quality in the term of signal-to-noise (SNR) ratio from 10.1 to 10.7 dB. The capability of improving MVDR’s performance has shown that the author’s proposed method can be integrated into a multi-channel system for solving more complicated problems.
In underwater communication networks, packet collisions at the receiving node, followed by transitions to a sleep state and subsequent retransmissions at inappropriate times, can lead to further collisions in subsequent transmission attempts. This will result in continuous interruptions in the transmission process, causing the communication network to become paralyzed. In this study, we propose an Automatic Repeat Request (ARQ) strategy for the Medium Access Control (MAC) protocol, applied to underwater communication channels, called the ARQ_WUF (ARQ Wake Up Factor) strategy. This approach reduces the number of retransmissions due to packet collisions at the receiving node, thus increasing transmission efficiency and effectively utilizing communication channel resources. In this strategy, when collisions occur, the nodes switch to a sleep state, and the wake-up time for retransmitting the packet is calculated to avoid packet collisions at the receiving node in the next transmission attempt. This will reduce the number of retransmissions caused by packet collisions at the receiving node, increase transmission efficiency, and make more effective use of communication channel resources.
Accurate and fast detection of tomato diseases is critical for good crop management and sustainable agriculture, as illnesses identified correctly and promptly can lead to early treatment of the plants. Due to the complexity and similarity of different tomato diseases and pests in the natural environment, there is a need for state-of-the-art object detection models to meet the demands for real-time and accurate detection. YOLOv8 and RT-DeTR are the most advanced object detection models, boasting great accuracy and real-time detection speeds. This research aims to compare the efficacy of the YOLOv8 model and the RT-DeTR model in tomato leaf disease detection. Our dataset is acquired using two methods: selecting from Kaggle a dataset containing a single tomato leaf in each image and capturing real-life photographs of tomato leaves in Da Lat City. Additionally, the dataset is augmented using several approaches, such as left and right rotation, blurring, and contrast adjustment. Various training strategies are utilized such as fine-tuning the pre-trained models using our dataset and experimenting with different optimizers, learning rates, weight decay value, and number of epochs. The numerical results show that RT-DeTR has a better precision score but detects significantly slower than YOLOv8. Moreover, YOLOv8’s inferred results make slightly less leaf detection and accurate classification than RT-DeTR, whose results also occasionally beat YOLOv8's detection confidence score.
Semantic communication (SC) has emerged as a novel paradigm aimed at revolutionizing the communication landscape through the efficient exchange of semantics. However, existing approaches are limited by redundant communication and the need for joint training at transceivers. To address these limitations, we propose a framework called Goal-oriented Invariant Representation-based SC (SC-GIR). The proposed method leverages contrastive learning to obtain an invariant and meaningful representation of source data that is task-agnostic. This latent representation facilitates efficient communication while retaining core features crucial for downstream task execution. Focusing on machine-to-machine tasks, we utilize a covariance-based contrastive learning approach to derive semantically dense latent representations. To evaluate the performance of the proposed approach, we consider different datasets for lossy compression. These compressed latent representations are subsequently employed in a goal-oriented artificial intelligence (AI) task. Experimental results demonstrate that the proposed SC-GIR framework significantly outperforms baseline schemes and conventional approaches.
This study investigates the physical layer secrecy of a backscatter communication (BC)-based multi-user time division multiple access (TDMA) mobile edge computing (MEC) Internet of Things (IoT) system. Specifically, an unmanned aerial vehicle (UAV) equipped with a server hovers over an energy-constrained mobile IoT device cluster to transfer RF energy and assist them in executing their tasks by employing the backscatter communication scheme. To enhance the secrecy of the considered system, a partial offloading scheme is proposed based on the non-linear harvested energy levels of users. Accordingly, a closed-form expression for the secrecy successful computation probability (SSCP) is derived, which jointly considers the harvested energy for local computing, offloading and processing latency, as well as secrecy capacity. The impact of key parameters on secrecy and latency performance is investigated to reveal the system’s behavior. Numerical results from computer simulations are provided to verify the correctness of our analysis.
The proliferation of the Internet-of-things (IoT) has led to an unprecedented increase in versatility of wireless applications. This evolution has imposed stringent and diverse quality-of-service (QoS) requirements on modern communication systems. These requirements complicate the design of wireless systems and increase resource consumption. The problem is exacerbated by the ever-increasing scale of IoT networks. This necessitates the development of green technologies, which enable energy-efficient operations. Reconfigurable intelligent surfaces (RIS) are emerging as a promising solution to the scalability problems of next-generation systems. RISs are composed of reconfigurable elements capable of altering the communication channel with minimal power consumption. Conventional RISs have independently functioning elements and are limited in their reconfigurability. To provide advanced beamforming capabilities, beyond diagonal RISss (BD-RISs) employ fully connected elements which increase reconfiguration precision. Motivated by the scale of next-generation networks, this article formulates and solves resource allocation problems, which jointly design the transmit beamforming vectors at the base-station and the reconfiguration matrix at the BD-RIS in a multi-user multiple-input single-output (MU-MISO) system. We utilize deep reinforcement learning (DRL) to solve these problem and evaluate the impact of various objective functions on the performance of the BD-RIS aided system.
The rapid advancement of technology has led to new applications, such as online gaming, virtual reality, and augmented reality, which demand substantial computing power. However, deploying these applications on mobile devices poses challenges due to their limited computing capacity. Mobile edge computing addresses this issue by enabling users to offload tasks to nearby edge servers. In this paper, we propose using unmanned aerial vehicles (UAVs) as mobile edge servers, leveraging their flexibility to establish high-quality user connections. UAVs equipped with robust computing resources can assist in processing tasks either partially or entirely, thereby facilitating the widespread adoption of modern applications. The objective of this study is to minimize overall network latency while ensuring quality of service, defined by minimum latency for user tasks and a guaranteed minimum data rate for offloading. We present an approximate solution method to address the offloading decision and power allocation problems. Simulation results demonstrate the superior performance of our approach compared to existing benchmarks.
We consider the downlink of a massive multiple-input multiple-output (mMIMO) system for cell-free networks where many dispersed Unmanned Aerial Vehicle-Satellite Communication (UAV-SATCOM) nodes (USCs) serve a few number of centric users under duplex operation with time division. A crucial concern while implementing cell-free mMIMO networks is the large amount of electricity consumed, which is proportional to the number of USCs. The topic is the usage of electricity green communications in terms of spectral efficiency has been brought up by this problem over multiobjective Quality of Service (QoS) per user. To address this issue, we create a brand-new low-complexity power management method with zero-forcing beamforming design to optimize the spectral efficiency of cell-free mMIMO taking into account the design of user-centric performance.
