Annals of Telecommunications最新文献

Emotion recognition (ER) is a pivotal discipline in the field of contemporary human–machine interaction. Its primary objective is to explore and advance theories, systems, and methodologies that can effectively recognize, comprehend, and interpret human emotions. This research investigates both unimodal and bimodal strategies for ER using advanced feature embeddings for audio and text data. We leverage pretrained models such as ImageBind for speech and RoBERTa, alongside traditional TF-IDF embeddings for text, to achieve accurate recognition of emotional states. A variety of machine learning (ML) and deep learning (DL) algorithms were implemented to evaluate their performance in speaker dependent (SD) and speaker independent (SI) scenarios. Additionally, three feature fusion methods, early fusion, majority voting fusion, and stacking ensemble fusion, were employed for the bimodal emotion recognition (BER) task. Extensive numerical simulations were conducted to systematically address the complexities and challenges associated with both unimodal and bimodal ER. Our most remarkable findings demonstrate an accuracy of (86.75%) in the SD scenario and (64.04%) in the SI scenario on the IEMOCAP database for the proposed BER system.

This paper, as an extended version of a communication presented at the ISIVC’2024 conference, deals with security issues in the software-defined networks (SDN); it introduces a Distributed Denial of Service (DDoS) detection system leveraging deep learning (DL) features. The main objective is to enhance SDN security by accurately classifying DDoS attacks, improving efficiency, particularly for zero-day attack detection, and enabling targeted mitigation strategies. Our contribution focuses on refining a hybrid DL model with a novel architecture that applies algorithms simultaneously to distinguish the normal SDN traffic and five carefully selected other classes covering various attack kinds, using an optimized CIC-DDoS2019 dataset for more efficient classification. Compared to the conference paper, the model has been reinforced by the use of attention mechanisms and transformer architectures in addition to layers’ adjustments and hyper-parameters re-settings. Additionally, the previously used training and testing data have been combined and split into three sets: 70% for training, 15% for validation (continuous partial evaluation), and 15% for final testing. The resulting solution (hybrid DNN-LSTM) demonstrated continuous exponential improvement of validation accuracy during the training step, recording a higher value near 99% and achieving a final testing accuracy of 98.84%. The improved model is suitable for real-world SDN systems, with its deployment, potential challenges, and practical benefits discussed.

This research paper introduces a reconfigurable and compact microstrip patch antenna designed and optimized for sub-6 GHz frequency bands, aligned with advancements toward millimeter-wave applications. The proposed antenna operates within the 2412–2484 MHz band for WiFi and the 3300–3800 MHz band for 5 G mobile phone communications. The antenna features a circular patch structure with compact dimensions, facilitating integration into miniature components and devices for wireless applications. To achieve frequency reconfigurability, a PIN diode was used as the switching technique. The antenna dimensions were optimized and simulated using CST and HFSS software. The simulation results were validated through measurements of the manufactured antenna. The antenna was fabricated on an FR4 epoxy substrate with a relative permittivity of 4.4 and a thickness of 1.6 mm.

Accurate prediction of mobility coordinates (x and y) is essential for effective transportation planning, urban development, and mobile network optimization. This study presents Tri-Sequence Temporal Network (TriSeqNet), an innovative architecture that synergizes the capabilities of bidirectional long short-term memory (BiLSTM), residual gated recurrent units (Residual GRU), and temporal convolutional networks (TCN) to concurrently predict x and y coordinates. Our approach outperforms existing methods by leveraging the combined strengths of these advanced neural network models. The performance of TriSeqNet is evaluated using traditional metrics such as mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE), as well as the coefficient of determination (R²) and explained variance (EV). This comprehensive evaluation framework demonstrates the robustness and accuracy of the proposed model in various predictive scenarios.

LoRaWAN provides extensive coverage, low energy consumption, and support for numerous connected devices. Aiming to reduce power demand while maximizing network throughput, LoRaWAN employs the ADR mechanism, which adjusts transmission parameters based on the link budget. However, standard ADR struggles in environments with mobile end devices and frequent signal variations, requiring alternative approaches for such scenarios. In this context, this paper proposes and evaluates Percentile-based ADR (P-ADR), a scheme that leverages statistical methods to estimate link conditions more accurately and swiftly adapt to dynamic environments. To assess its performance, P-ADR was compared against ADR+, M-ADR, and standard ADR in various urban and suburban scenarios, considering simulated networks with 1–2 gateways and 200–1000 static and mobile end devices. Results show that P-ADR significantly enhances performance in mobile environments, achieving up to a 22.6% improvement in the average PDR and up to 62.63 bits/J higher average energy efficiency compared to standard ADR. These findings suggest that P-ADR is a promising solution for IoT applications, particularly in scenarios with fluctuating channel conditions.

Federated learning offers a promising solution for enabling collaborative model training across autonomous vehicles while preserving privacy and reducing communication overhead. However, efficiently selecting clients for the training process remains challenging, particularly in environments with statistical heterogeneity and frequent client failures. Client failures, often due to mobility or resource constraints, can significantly degrade the performance of the global model by reducing accuracy, slowing convergence, and introducing bias. This paper proposes a novel approach to enhance the robustness and reliability of FL in autonomous vehicle networks by integrating an entropy-based client selection mechanism with a minimal repair model. The entropy-based selection identifies clients with diverse and informative data, while the proposed tool substitutes failed clients with similar ones using the Hausdorff distance. Our results demonstrate that this combined approach outperforms existing methods regarding training loss, accuracy, and area under the curve, particularly in scenarios with high client dropout rates. These findings highlight the importance of considering data diversity and client substitution strategies to maintain robust FL in dynamic vehicular environments.

In order to improve the security of cognitive non-orthogonal multiple access (NOMA) systems, we propose a highly secure forwarding strategy for secondary networks, in which the relay uses the data sent by the primary transmitter to improve the security of its forwarding data. Specifically, two relay encoding strategies are designed to improve the security of the cognitive transmitter data, namely, power superposition (PS) encoding strategy and bit-level exclusive OR-PS (XOR-PS) encoding strategy. Considering the imperfect successive interference cancellation (SIC) technology, the exact closed-form expressions of the outage probabilities and intercept probabilities of the PS and XOR-PS schemes in Rayleigh fading scenarios are derived. Then, the corresponding approximate results in high signal-to-noise ratio (SNR) are also given and the correctness of the theoretical results is verified by simulations. Furthermore, two other conventional PS and XOR-PS schemes without using the data of the primary user, represented by NPS and NXOR-PS, respectively, are provided as the benchmark to compare the security and reliability of the proposed protocols. Finally, numerical results show that the security of the proposed PS and XOR-PS schemes is much better than that of the NPS and NXOR-PS schemes in the case of without degrading the outage performance.

The blockchain trilemma, involving the balance of scalability, security, and decentralization, remains a critical challenge in blockchain networks. Layer 2 (L2) solutions, especially rollups, have emerged as promising approaches to enhancing scalability. They execute transactions on an auxiliary L2 chain and submit batched results to the Layer 1 (L1) blockchain, preserving L1 security and decentralization while significantly increasing throughput and reducing transaction costs. However, current rollup mechanisms primarily focus on batching and compression techniques without dynamically optimizing transaction execution based on resource utilization. This paper extends our previous work on AI-driven opcode smart contract classification, presented at the ISIVC2024 conference [1], by introducing a new layer of optimization through an opcode-based adaptive rollup execution strategy. By analyzing opcode sequences of smart contract transactions, we categorize transactions based on computational complexity and resource requirements. Our adaptive batching algorithm prioritizes transactions using an opcode-based score, forming batches that optimize gas consumption, enhance throughput, and improve processing efficiency within rollup mechanisms. Additionally, we incorporate dynamic scheduling algorithms within the sequencer, utilizing machine learning models to predict optimal execution orders and adjust strategies based on real-time network conditions. Our analysis evaluates the performance of our adaptive batching algorithm against traditional methods and assesses the dynamic scheduling approach as an enhancement to our model. The results indicate improvements in sequencer efficiency and resource utilization during rollup transaction execution. This research contributes to addressing the blockchain scalability trilemma by offering an adaptive approach that responds to evolving blockchain network demands.

The healthcare sector has witnessed a transformative shift in recent years, driven by rapid advancements in digital technologies. Among the myriad of applications, the management of Alzheimer’s disease (AD) has garnered significant attention. AD, the most common form of dementia, affects millions globally and presents a significant challenge due to its progressive and currently incurable nature. Early detection is crucial, yet existing diagnostic methods are invasive, expensive, and not readily accessible. This study proposes a hybrid approach combining traditional acoustic features (e.g., MFCC, pitch, jitter, shimmer) with deep learning-based embeddings (YAMNet, VGGish) to enhance the robustness and accuracy of AD detection through speech analysis. The methodology involves comprehensive feature extraction, dimensionality reduction via autoencoders, and classification using advanced machine learning (ML) and deep learning (DL) models. Evaluation on the ADReSS dataset demonstrates the proposed method’s superior performance, achieving an accuracy of 89.9% with a deep neural network classifier. The results highlight the potential of integrating traditional and modern techniques to develop non-invasive, cost-effective, and accessible tools for early AD detection, paving the way for timely intervention and improved patient outcomes. Future work will focus on expanding datasets, incorporating diverse demographics, and refining models for better sensitivity and specificity in clinical applications.

Detecting offensive speech poses a challenge due to the absence of a universally accepted definition delineating its boundaries. However, the scarcity of labeled data often poses a significant challenge for training robust offensive speech detection models. In this paper, we propose an approach to handle data scarcity through data augmentation techniques tailored for offensive speech detection tasks. By augmenting the existing labeled data with speech samples generated through noise injection, our method effectively expands the training dataset, enabling more comprehensive model training. We evaluate our approach on Vera Am Mittag (VAM) corpus and demonstrate significant improvements in offensive speech detection performance compared to that without data augmentation. Our findings highlight the efficacy of data augmentation in mitigating data scarcity challenges and enhancing the reliability of offensive speech detection systems in a real-world scenario.