Implementing shortest-path algorithms in traditional IoT networks is challenging due to their importance in efficient routing and data transmission. In conventional networking, determining the shortest path is typically handled by distributed devices, such as routers and switches. However, this decentralized nature often leads to suboptimal routing decisions and inefficient resource use. As a result, network performance suffers, affecting service quality and increasing latency. Software-Defined Networking (SDN) offers a promising solution to address this challenge due to its centralized control, flexibility, and programmability. In this work, we integrate the Dijkstra algorithm within an SDN environment using the Ryu controller to enhance routing efficiency. The paper presents a detailed performance analysis, focusing on critical metrics such as Throughput, Jitter, Data Transmission Rate, and Packet Loss. Through extensive simulations and practical implementation, the impact of Dijkstra-based routing is compared against the default routing strategy. The observed results demonstrate improved data flow control, reduced packet loss, and consistent transmission rates, indicating better adaptability and resilience. This study reinforces the effectiveness of combining Dijkstra’s algorithm with SDN to optimize network operations and service delivery in modern IoT networks.

Recent breakthroughs in Natural Language Processing (NLP) have caused a surge in the application of Large Language Models (LLMs) for a variety of tasks, including content generation, translation, and decision support. However, these models tend to re-represent and magnify human biases present in their training data, leading to socially unfair predictions. We study bias mitigation through fine-tuning and prompt manipulation in three open-source models— LLaMA2-7B, Mistral-7B, and Dolly-7B. We develop six different prompting schemes and generate an inference dataset to measure how they mitigate bias. We also fine-tune every model to encourage neutral representations with respect to four social factors, including gender, race, occupation, and religion. Post-finetuning evaluations show that the neutrality and fairness of model outputs are greatly improved. The findings demonstrate the interplay of prompt variation and fine-tuning in mitigating societal bias in LLMs, guiding the building of more fair and ethical language technologies.

The escalating threat of multimodal misinformation, propagated through manipulated text and imagery, necessitates robust and resilient detection frameworks. As generative AI advances, adversarial attacks exploit both modalities to bypass existing defenses, posing a critical challenge. While multimodal fake news detection models offer promising solutions by integrating textual and visual features, their inherent vulnerabilities under adversarial conditions and their behavior in incomplete modality scenarios remain insufficiently explored. This study provides a systematic evaluation of three widely-used multimodal architectures−BERT-VGG, CNN-VGG, and LSTM-VGG−on two benchmark datasets, D1 and RECOVERY. We significantly extend the scope of robustness assessment by subjecting these models to a diverse range of adversarial attacks, including ten image-based and five text-based perturbations, alongside novel large language model (LLM)-driven manipulations using LLAMA3. Furthermore, our research rigorously investigates the impact of modality removal (unimodal settings) on detection performance, offering crucial insights into the intrinsic reliance on specific modalities. Our findings reveal significant vulnerabilities within current multimodal fake news detection models, particularly their susceptibility to advanced text-based adversarial attacks and evidence of “Unimodal Bias." Crucially, the image modality is demonstrated to play an indispensable role in maintaining detection efficacy and resilience. This comprehensive analysis highlights the urgent need for adaptive strategies and advanced architectures that can withstand sophisticated adversarial tactics and manage modality inconsistencies, paving the way for the development of more resilient and effective next-generation misinformation detection systems.

The proliferation of sophisticated multimodal disinformation, fueled by advances in generative AI and cross-modal manipulation, presents escalating challenges for existing Multimodal Fake News Detection (MFND) systems. In this work, we present a comprehensive and rigorous evaluation of the robustness spectrum of MFND architectures. We analyze three canonical architectures (LongBERT-VGG, CNN-VGG, LSTM-VGG) alongside two recent State-of-the-Art (SOTA) models, MFIR (2023) and CroMe (2025), across three heterogeneous benchmark datasets: D1, RECOVERY, and PolitiFact. These datasets encompass a wide range of content modalities and class distributions, enabling a robust assessment of model generalizability. We go beyond conventional evaluation paradigms by subjecting these architectures to a broad suite of adversarial perturbations, including ten image-based and seven text-based attacks. Notably, we introduce a novel class of instruction-tuned Large Language Model (LLM)-generated adversarial texts and evaluate their effectiveness using state-of-the-art models, Meta’s LLaMA 3 8B and DeepSeek-LLM 7B Chat. Our findings reveal a consistent degradation in model performance under these advanced textual manipulations, with DeepSeek-based perturbations demonstrating marginally higher efficacy in evading detection. Further, we perform a detailed modality ablation study on both foundational and SOTA models. Results indicate that while modern architectures like CroMe improve baseline fusion through metric learning, they remain critically vulnerable to semantic inversions that preserve syntax but reverse veracity. This study offers a foundational benchmark for future research in adversarially robust MFND, providing actionable insights into modality-specific fragilities and the evolving threat landscape of AI-driven misinformation.

The increasing sophistication of multimodal disinformation driven by generative AI and cross-modal manipulation poses significant challenges to the reliability of existing Multimodal fake news detection (MFND) systems. This work presents a rigorous interrogation and comprehensive evaluation of canonical MFND architectures, rather than proposing a new system. We assess the feasibility and robustness of five diverse architectures, LongBERT-VGG, CNN-VGG, LSTM-VGG, BERT-CLIP, and the end-to-end transformer ViLBERT across three benchmark datasets: D1, RECOVERY, and PolitiFact, which collectively represent varying degrees of class balance and content characteristics. Our analysis transcends conventional evaluations by subjecting these models to an extensive battery of adversarial perturbations: ten distinct image-based attacks and seven sophisticated text-based attacks. Crucially, we introduce novel adversarial textual manipulations generated via instruction-tuned LLMs, specifically comparing the potency of attacks derived from Meta’s Llama 3 8B against those from DeepSeek-LLM 7B Chat. We further introduce an analysis of black-box transfer attacks to simulate real-world threats. Furthermore, we systematically investigate the performance implications of modality ablation, providing granular insights into the functional interdependence of textual and visual features. Our empirical findings, now reported as mean ± standard deviation over multiple seeds and validated with statistical significance tests, illuminate significant vulnerabilities inherent in contemporary MFND models, particularly when confronted with advanced LLM-driven textual attacks. Notably, our results indicate that adversarial text generated by the DeepSeek model exhibits a statistically significant superior capacity to degrade detection performance compared to Llama 3 8B. The analysis further underscores the pivotal, albeit context-dependent, role of the visual modality. We supplement our quantitative findings with a qualitative error analysis using Grad-CAM, revealing model failure modes, including over-reliance on dataset artifacts. We also provide a practical analysis of computational costs and potential defense strategies. By advancing the frontiers of adversarial resilience and modality dependence analysis in MFND, this work establishes a foundational benchmark for the conception and validation of next-generation disinformation detection systems.

Intelligent cities use Vehicular Adhoc Networks (VANETs) to ensure transport safety and prevent accidents in real-time. A stable network depends on the efficient cluster and Cluster Head (CH) election because of high variability and unpredictable drive behaviour. The present investigation introduces a CH selection strategy that improves transmission efficiency in congested metropolitan areas by utilising Adaptive Neighbouring Behaviour-Incentivization (ANB-I). Conventional graph-based clusters result in diminished connection dependability and interruptions. It culminates in decreased cost and latency. When it comes to CH selection, the ANB-I system takes transport behaviour into account, giving more weight to CHs that have more consistent trends and greater connection times. Its objective is to compare the different incentivised parameters to enhance VANET efficiency over a range of vehicle volumes; it has been evaluated through MATLAB and SUMO tools. The study analyses involve 100 to 800 vehicles for implementing the ANB-I. Also represents the comparison between two selected clusters based on the adaptive behaviour.

Large Language Models (LLMs) are increasingly being applied to sensitive domains such as medical care, which pose multiple technical and ethical challenges. The most immediate issues include biases built into these models, which are trained on large datasets that may reflect societal prejudices. Such biases can yield outputs that are unfair, ungeneralizable, or even harmful, rendering them clinically inapplicable in the real world and noncompliant with ethical and regulatory constraints. We examine how well bias suppression works when fine-tuned models are prompted across four critical categories (gender, race, profession, and religion). We manually curate a diverse inference dataset and create twelve prompt variants -- six of which are debiased -to test the outputs of three open-source LLMs: Llama2-7B, Mistral-7B, and Dolly-7B. The fairness and interpretability of the outputs are evaluated using a bias-scoring metric, where lower scores indicate better fairness and interpretability. We also note that debiased prompts reduce bias, and fine-tuning the model performs even better. The results emphasize the critical importance of timely action, model robustness, and ongoing ethical scrutiny for trustworthy and fair LLM deployment in real-world settings, such as medical imaging, and the maintenance of Electronic Health Records (EHR).

Recognition of human activity is a challenging issue, especially in the presence of multiple actions and multiple scenarios. Therefore, in this paper, multi-view multi-modal based human action recognition (HAR) is proposed. Here, initially, motion representation of each image such as Depth motion maps, motion history images, and skeleton images are created from depth, RGB, and skeleton data of RGB-D sensor. After the motion representation, each motion is separately trained by using a 5-stack convolution neural network (5S-CNN). To enhance the recognition rate and accuracy, the skeleton representation is trained using a hybrid 5S-CNN and Bi-LSTM classifier. Then, decision-level fusion is applied to fuse the score value of three motions. Finally, based on the fusion value, the activity of humans is identified. To estimate the efficiency of the suggested 5S-CNN with the Bi-LSTM method, we conduct our experiments using UTD-MHAD. Results show that the suggested HAR method attained better than other existing approaches.

In today's dynamic networking landscape, integrating Software-Defined Networking (SDN) with Traffic-Expert Virtual Networks (TEVN) presents a promising avenue for optimizing network performance. This research investigates the implementation of TEVN Embedding within SDN frameworks, utilizing the Ryu controller to address inefficiencies in traditional virtual network embedding algorithms. Methodologically, the study proposes a framework for TEVN and evaluates its performance against benchmark methods using various parameters such as throughput, bandwidth, packet loss, and Round-Trip Time (RTT). The evaluation is conducted through extensive experimentation in simulated SDN environments, with results analyzed and compared comprehensively. The findings reveal that TEVN significantly improves network efficiency, achieving higher throughput, lower latency, and reduced packet loss compared to default embedding algorithms. These results underscore the potential of TEVN to revolutionize network management practices, offering a promising solution for addressing the evolving challenges of modern network infrastructures. This research contributes to advancing SDN technologies and gives insights into enhancing network efficiency in dynamic environments.

There are significant challenges in the domain of sign language machine translation such as dataset quality, translation system complexity, and linguistic distance measurement. We introduce SignPro, a refined version of the SignBank dataset tailored specifically for machine translation purposes. We also address the challenge of comparing sign language translation methods across publications by providing an implementation of a text to gloss translation pipeline with grammar inclined word reordering and neural machine translation.

The rapid growth of connected vehicles in vehicular ad-hoc networks (VANETs), with frequent data exchanges between vehicles and roadside units (RSUs) enhances traffic efficiency but also increases the attack surface, as adversaries can exploit the high communication density to launch malicious activities. Specifically, attacks such as denial-of-service (DoS), where attackers flood the RSU with excessive packets or fake requests, under dense traffic conditions where each vehicle already transmits periodic safety messages, this artificial load quickly overwhelms the RSU processing capacity, leading to congestion and service disruption. To mitigate such risks, intrusion detection systems (IDSs) are essential, as they monitor network traffic in real time, identify malicious patterns, and prevent attacks before they compromise critical services. Existing deep learning-based IDSs algorithm often treat these threats in isolation, overlooking the semantic relationships among different attack types. To address this gap, we propose a novel knowledge graph-based deep convolutional neural network (KG-DeepCNN) IDS for securing VANETs. The proposed IDS is deployed at the RSU level (RSU-IDS), where incoming vehicle and packet information are analyzed to distinguish legitimate requests from malicious ones, thereby strengthening vehicle-to-vehicle (V2V) communication security. The knowledge graph enriches raw data by capturing semantic correlations among attack patterns, which are then processed by an ensemble of CNN learners with diverse architectures for improved classification accuracy. The framework is evaluated on four benchmark datasets—BoT-IoT, ToN-IoT, UNSW-NB15, and NSL-KDD—achieving accuracies of 99

Vehicular Ad hoc Networks (VANETs) are gaining popularity among academic institutions and industrial fields. The research area of VANET captivates a significant amount of curiosity from authors across the globe. Nonetheless, strong scientific contributions have been presumed through VANETs to explore the clustering concept with VANETs. For this reason, it is essential to outline the current state of study in this discipline. Since 2010, the number of research articles has grown exponentially; the assessment through a tool is needed for authors to comprehend the existing findings and conclusions throughout this field. Using different factors, this article analyzes the social standing of research developments in the discipline of “Clustering in Vehicular Ad Hoc Networks” from 2010 to October 2024 with the help of bibliometric analysis. Since 2024, 17,461 publications have been connected well with the VANETs realm, which was retrieved first from the Scopus database (in BibTeX format). Furthermore, 377 of the collection's articles pertain to the clustering techniques within VANETs. To achieve the best result, the bibliometric tool ‘R-Tool’ using ‘Biblioshiny’ generates and visualizes the specialized area's bibliometric connectivity. Also, this article would provide a visual representation of the scientific data gathered from various authors, countries, and connections to assist authors in identifying the clustering field in the VANETs.

Face recognition processes often take into account the orientation and illumination of the face to achieve a satisfactory recognition rate. Additionally, extracting features aimed at capturing expressions is crucial for expression recognition. By understanding emotions through facial and expression recognition, it becomes possible to monitor and manage office environments more effectively, aiding in people management. This research aims to identify specific features, including facial orientation, to enhance the recognition rate. The cutting-edge solution incorporates facial landmarks, affine transformation, and advanced feature extraction techniques utilizing the Attention module and Local Binary Pattern (LBP), addressing issues related to face orientation, illumination, and targeted features. This paper introduces an improved face recognition system that employs a preprocessing layer incorporating facial landmarks and affine transformation, a modified attention module, and circular derivative LBP within the present state-of-the-art framework. The proposed model also includes new data augmentation methods, such as shear and zoom, augmenting the base method. The system’s evaluation spans four datasets: MMA, JAFFE, FER2013, and CKplus. The proposed system achieves an average accuracy of 95.63

Predicting blood glucose is highly significant for patients with diabetes to manage their condition efficiently. Deep learning (DL) approaches have demonstrated great potential in blood glucose prediction modeling. By leveraging time-series data from continuous glucose monitoring (CGM) devices, this model can capture complicated temporal dependencies and patterns in glucose dynamics. DL based blood glucose prediction models learn from insulin dosages, past glucose readings, physical activity levels, meal intake, and other related features to forecast future blood glucose levels with maximum accuracy. Thus, this study presents an Optimal Attention-based Long Short-Term Memory for Blood Glucose Level Prediction (OALSTM BGLP) model. Firstly, it employs Min-Max scaling to normalize the input data, ensuring consistent and meaningful comparisons across different features. Additionally, the model generates time series data for multiple forecasting horizons, including 15, 30, 45, Aand 60-minute(m) intervals, enabling flexible and dynamic predictions to accommodate various planning and decision-making needs. Moreover, the OALSTMBGLP technique uses the ALSTM model, which incorporates an attention mechanism to selectively concentrate on relevant information within the input sequence while capturing long-term dependencies. This attention mechanism permits the method to effectively extract salient features from the input data, enhancing its predictive capabilities. Furthermore, the model is optimized using the RMSProp optimizer, which adjusts the rate of learning dependent on the magnitude of recent gradients, facilitating efficient training and convergence. The performance evaluation of the developed technique on the OhioT1DM dataset shows its promising performance over recent state-of-the-art methods.

Language models (LMs) like RoBERTa are known to have social biases to varying degrees. These biases appear along various social lines (eg, gender, race, religion, occupation), and in broader sociocultural contexts that perpetuate stereotypes with potential ethical implications. In this paper, we systematically investigate several debiasing strategies on the CrowS-Pairs benchmark dataset to evaluate the extent to which stereotype-versus-anti-stereotype preferences manifest in the model and how they can be mitigated. We measure against various pre-processing, in-processing, and post-processing techniques, including Counterfactual Data Augmentation (CDA), Counterfactual Data Substitution (CDS), fine-tuning, adversarial debiasing, and post-hoc calibration. Performances reveal that advances in in-processing methods, especially adversarial training with LoRA and post-hoc calibration (bias score of 52.59), achieve high-level accuracies compared to the baseline RoBERTa-large (bias score of 32.63), illustrating the benefits of hybrid debiasing pipelines for fairness in LMs. Achieving a bias score of 50 is ideal. We also investigate the extent to which effective debiasing techniques can be improved without compromising language fluency.

Social media fake news evolves rapidly, rendering static detection models ineffective. FakeNews Buster is a dynamic continual learning framework addressing concept drift in misinformation. It integrates BERT and the lightweight FastFormer, using fixed-length and recurrent processing for varied text formats like tweets. A custom loss function manages uncertainty in noisy social data, while knowledge distillation retains critical information. An explainability module combining the Google Fact Check API with vector-based semantic search enhances interpretability, providing evidence-based justifications. We explore multiple training strategies and assess performance using standard classification metrics. The model’s vocabulary is expanded with domain-specific terms and emojis to reflect informal online content. Results show FakeNews Buster offers improved adaptability, transparency, and trustworthiness in fake news detection through its novel integration of continual learning with external fact-checking capabilities.

The monitoring of blood glucose levels (BGL) in the human body is a successful step in the fight against diabetes. Continuous glucose monitoring (CGM) devices are designed to monitor BGL but are unable to detect either lower or higher insulin levels in the future. These situations resulted in the occurrence of hypoglycemia and hyperglycemia. Predicting BGL accurately is a challenging task because of the nature of its time series data. Adverse diabetes conditions can only be avoided if the BGL has been monitored correctly. The existing mathematical models have high inter-patient variability, which does not allow such methods to be applied in clinical settings. This has led to the search for an artificial intelligence-based solution. Machine learning-based methods suffer from significant feature extraction problems in time series data. The proposed novel method utilises a 1D convolution layer (Conv1D) to extract deep hierarchical features and a long short-term memory (LSTM) recurrent connection to capture crucial temporal information. The framework uses CGM data for a 60-minute prediction horizon (PH). The experiment results demonstrate that the proposed framework, which utilises a convolutional recurrent connection, achieved superior results compared to other recent state-of-the-art models.

Although transformers have achieved remarkable performance across different vision tasks, they have not yet demonstrated the comparable proficiency to CNN (ConvNets) in image generation. This paper presents an innovative Swin Transformer plus VQVAE model for encoding and decoding. Elevating Swin Transformer’s scalability and global dependency modeling, our proposed model addresses these shortcomings. This advancement fills the gap in transformer-based solutions engineered specifically for image generation thereby offering improved quality and efficiency in generating images.

This research focuses on addressing the challenges of visual emotion analysis and emphasizes the need for improved comprehension and categorization of emotions by machines. The objective is to develop an efficient architecture for emotion classification in real-world scenarios encountered in multimedia retrieval tasks, considering factors like illumination, occlusion, pose variations, small face sizes, multimodal detection, and big data issues. The proposed Dense Blocked Network-based VEDANet architecture surpasses the state-of-the-art on benchmark datasets by leveraging pre-trained CNN architectures to recognize facial features and extract metadata. The exploration of 128 descriptors from a deep residual network further enhances the operations. Using the VEDANet framework, emotions are classified into seven categories with superior accuracy compared to conventional approaches. The study investigates the efficacy of an over-the-top optimization (OTO) layer to enhance emotion classification. The proposed model achieves impressive accuracy scores on diverse datasets, including AffectNet, Google FEC, Yale Face DB, and FER2013, with percentages of 87.30

Vehicle ad hoc networks (VANETs) have garnered considerable attention for their potential to enhance road safety and facilitate advanced driver assistance systems. A fundamental aspect of VANET is the formation of stable clusters and cluster heads (CH) for improved network performance. Due to the dynamic nature of VANET and the different mobility of the vehicles, maintaining CH stability is significant. To address this issue, this study presents the network as connecting hypergraphs. The proposed approach uses an improved tensor-trace maximization-based spectral clustering algorithm (iTTM) and eigen heuristics to generate an optimal. The suggested clustering approach is followed by the CH selection based upon the four vehicle attributes: modularized link lifetime, connectivity level, relative speed, and consensus trust score. The multi-decision CRITIC approach will decide the CH in each cluster using those four attributes. These metrics have improved throughput and reduced packet delay to improve network performance. Simulation results, conducted in Delhi's Connaught Place using SUMO, demonstrate the proposed method's superiority in CH stability, throughput, and packet delay compared to existing algorithms. The state-of-the-art comparison is done on the criteria of CH stability, which comes out to be 84