In real life, anomalies are usually rare and highly different from the majority class. This makes it challenging for traditional approaches to discover them without biasing the results or overlooking little patterns. Finding anomalies in attributed networks is still hard, especially when there aren’t enough of them, and they aren’t evenly spread out among the majority class. We present CAFA (Class-Aware Feature Attention), a new way to improve detecting anomalies in large-scale attributed networks. We combine class-aware feature attention with Graph Convolutional Networks (GCNs) in our method to better find unusual nodes by dynamically focusing on the most important attributes. We start by normalising the features in the dataset and using K-Means to assign them artificial labels when there aren’t any actual ones. The feature attention layer is the most important module of CAFA. It learns to give more weight to features that are most important for finding anomalies, which makes it easier to find them. After training, we use dJSD to look at the structural features of the graph and Euclidean Distance (ED) to look at the node attributes. Then we combine these two scores to get a final anomaly score. Experimental results show that CAFA greatly improves the ability to find anomalies, especially rare ones, on benchmark datasets including Citeseer, Amazon-Photo, and Amazon Computers.

The fast growth of deepfake technologies poses a huge dilemma on digital content legitimacy, hence threatening public trust, privacy, and security. This paper introduces a hybrid deepfake detection model that utilizes EfficientNetB0, a lightweight convolutional neural network, with Gated Recurrent Unit (GRU) layers to leverage both spatial and temporal features of videos. The EfficientNetB0-based feature extraction combined with GRU-based sequential learning helps the model to handle inconsistencies between the successive frames in a video. The proposed model utilizes multi-scale feature extraction with transfer learning to capture the complex temporal and spatial patterns suggestive of video manipulation. Extensive experiments conducted demonstrating this as the state of the art were carried out on benchmark datasets, Celeb-DF and DFDC had an accuracy of 89.71

The rapid expansion of image editing tools has significantly complicated the verification of digital image authenticity in forensic, legal, and media environments. In practical scenarios, images frequently undergo multiple sequential processing steps, such as Gaussian blurring (GB), median filtering (MF), resampling (RS), additive white Gaussian noise (AWGN), and JPEG compression. These cascading operations create complex manipulation chains that partially conceal or erase forensic evidence from earlier phases, making detection particularly challenging in low-resolution and greyscale contexts. While recent convolutional neural network (CNN) techniques focus on local artefacts, they struggle to model long-range dependencies. Conversely, Transformer-based models often overlook the subtle noise patterns induced by localized manipulations. This research presents DFCT-Net, a Dense-Frequency-Context Transformer Network developed for robust multi-class chain-of-manipulation detection. DFCT-Net incorporates a residual high-pass enhancement module, a dense convolutional encoder for precise spatial forensic feature extraction, a patch-based Transformer for comprehensive contextual reasoning, and a lightweight frequency-domain energy branch to explicitly identify spectral distortions generated by filtering and resampling. A unified fusion technique effectively integrates this spatial, contextual, and frequency-sensitive information. Evaluated on datasets for both single-operation and multi-chain manipulations, DFCT-Net achieves a test accuracy of 97.73% in single-chain operations and 93.22% in multi-class classification. The results demonstrate that multi-domain feature fusion provides a highly robust and chain-aware approach to image forensic analysis.

The Internet of Things (IoT) environments are always changing and have a lot of different types of devices, which makes it hard to find anomalies because the way devices connect to each other and how traffic flows changes all the time. Traditional methods either overlook relational structure or inadequately model temporal change, hence constraining their robustness in practical applications. To address these limitations, we propose DySTAL-AD, a Dynamic Self-Attention and LSTM-based framework for binary anomaly detection in evolving IoT networks. The proposed method represents IoT traffic as a sequence of dynamic graph snapshots constructed using sliding time windows. Structural and temporal self-attention mechanisms jointly capture topology-aware dependencies and interaction drift, while an LSTM refinement layer enhances long-term behavioural modelling and anomaly persistence detection. Extensive experiments on CICIDS-2017, CICIDS-2018, and CICIoT-2022 datasets demonstrate that DySTAL-AD consistently achieves superior or highly competitive Precision, Recall, F1-score, and Accuracy compared to spatial-only (GCN, GAT), temporal-only (GRU), and dynamic attention-based (DySAT) baselines. The results validate that integrating dynamic graph attention with recurrent memory refinement improves detection robustness across both enterprise-scale and heterogeneous IoT environments. DySTAL-AD thus provides an effective and scalable solution for anomaly detection in dynamic communication systems.

The class imbalance problem is a prevalent challenge in the domain of network science for node classification tasks, often leading to biased models and poor performance on minority classes. To address this issue, we propose GCN-SynDCL, a model for imbalanced node classification that utilizes Graph Convolutional Networks (GCN), Multi-Layer Perceptrons (MLP), and synthetic node generation with Dynamic Curriculum Learning (DCL). The feature extraction is performed using GCN as an encoder and MLP as a decoder, where GCN captures both topological and feature-level information, while the MLP reconstructs node features to ensure semantic consistency. These embeddings are then used to generate synthetic nodes for minority classes, which are progressively integrated into training via DCL to improve representation and reduce overfitting. Next, synthetic nodes are integrated into the graph using adaptive k-nearest neighbor (k-NN) with cosine similarity to form structurally coherent edges. Finally, a GraphSAGE classifier is trained on the augmented graph, leveraging both original and synthetic nodes for improved classification performance across all classes, particularly the minority ones. Furthermore, comprehensive studies on benchmark datasets demonstrate that GCN-SynDCL consistently outperforms contemporary methods in handling the class imbalance problem. Our approach offers a flexible and robust solution to imbalanced node classification with broad applicability.

Skeleton-based methods in Human Activity Recognition have gained significant attention due to their critical applications in analyzing activities and interpreting human behavior. Despite advancement in the field, challenges such as incomplete skeleton data, often caused by occlusion, remain unresolved. These issues significantly degrade the performance of existing models, which mainly depend on complete skeleton data. To address these limitations, we propose a framework using Multi-stream Spatial Temporal Graph Convolution Networks (STGCNs) to identify activities in partially occluded environments. The proposed method integrates pose encoding and multimodal feature generation techniques to extract discriminative features from joints, even in challenging scenarios. A key innovation is the activation-based multi-stream strategy, where under-activated key joints are passed to subsequent streams for further feature extraction, ensuring a comprehensive understanding of human activities. To accurately identify the activities and ensure resilience against occlusion effects, outputs from all streams are then combined and sent to a Long Short-Term Memory (LSTM). The extensive experimental results estimate the accuracy of the proposed model, achieving state-of-the-art performance on standard datasets and significantly qualifying the effects of occlusion in synthesized datasets.

The understanding of time dependent dynamics in evolving network interactions is crucial for applications ranging across various domains. In this paper, we introduce TA-GC-LSTM (Temporal Adaptive Graph Convolutional Long Short-Term Memory) which uses deep learning framework with novel combination of models. This proposed model of ours, efficiently captures spatial dependencies through graph convolution, temporal sequences using LSTM, and gives selective importance to influential time steps through the attention mechanism. In contravention to traditional methods, which rely on static graph representations, TA-GC-LSTM dynamically learns the temporal evolution of node relationships, enhancing predictive accuracy in link prediction tasks. In our framework, we have carried out processing of datasets by binning interactions into fixed time windows, encoding unique nodes with learnable embeddings, and filtering sparse time steps to optimize computational efficiency. To validate our approach, we have tested the model on three real-world datasets and compared our model performance against Graph Convolution Embedded LSTM (GC-LSTM) and Temporal Graph Convolutional Network (T-GCN) as benchmarks across multiple evaluation metrics. Our results demonstrated that TA-GC-LSTM outperforms baseline models, achieving an AUC score of 93%, while maintaining computational efficiency, making it a robust solution for modelling evolving graph structures.

Activity recognition (AR) using surveillance cameras remains challenging because of the variation between the data required for deep learning and the capabilities of CCTV-based frameworks for data recording. Most research on skeleton-based activity focuses on improving recognition using complete data, often overlooking performance on incomplete skeleton data due to occlusion or noise. However, in practical situations, it is inevitable to obtain a noisy or occluded human skeleton. When some informative keyjoints are unavailable or disturbed, it can compromise the efficiency and effectiveness of existing methods. This article proposes a skeleton-based model for activity recognition to deal with the occlusion problem. The main approach of the proposed model is to address occlusion by treating it as a missing value imputation problem within the feature matrix. First, it extracts features such as normalized joints, relative joint coordinates, angles, and velocity using a CNN-based technique and encodes the skeleton data into a feature matrix. The innovative use of generative adversarial imputation networks (GAIN) with bidirectional long short-term memory (Bi-LSTM), enhanced by spatial and temporal attention mechanisms, effectively handled missing data due to occlusions, ensuring a complete feature matrix for subsequent classification. Additionally, this article introduces autoencoder-based Bi-LSTM with attention modules that improve activity classification by capturing and leveraging spatial and temporal dependencies. A comprehensive experiment was conducted on both occluded and unoccluded datasets to validate the effectiveness and improvements of the proposed model. Comparative results show that the proposed model outperforms the existing state-of-the-art methods, achieving superior recognition accuracy in an occluded environment.

Link prediction (LP) is a prominent research topic in network science and complex network analysis, focused on predicting future connections between unconnected node pairs by examining network topology and related characteristics. This paper introduces an advanced LP model combining the graph attention network (GAT) with Jaccard similarity and node centrality measures, such as local interaction density (LID) and hubs and authority centrality (HAC). Initially, a weight matrix incorporating structural and topological information is created using a feature matrix derived from similarity measures and node centrality. Then, node embeddings are generated using GAT, allowing the model to learn detailed representations that consider local and global contexts. GAT's multi-head attention mechanism enables the model to focus on various aspects of the node neighbourhood, capturing diverse structural information. A well-defined dataset is created from these embeddings, representing nodes at the endpoints of edges labelled as positive or negative. Linear discriminant analysis (LDA) is applied to this well-balanced and labelled dataset to perform classification tasks, leading to accurate link prediction. The model's performance is evaluated across eight different datasets, and the obtained results reveal the proposed model's superiority over several baseline and recently proposed LP models.

Ensuring the reliability of visual media has become more challenging due to the widespread dissemination of altered images on digital platforms. When photos are compressed or scaled, the forensic traces produced by methods such as median filtering, Gaussian blurring, resampling, or additive noise often become diminished or hidden, making the task even more difficult. Traditional picture forensics methods are negatively impacted in these situations, particularly when JPEG compression hides minute pixel-level artefacts that are crucial for precise detection. Introduce PRILNet, a compact hybrid architecture designed specifically for compressed image alteration detection, to address these limitations. The core of the model is the PixelRelative Intensity Layer (PRIL), which produces manually constructed descriptors that highlight subtle variations in local pixel patterns. By processing these descriptors through a hierarchy of convolutional stages enhanced with CBAM attention, the network can highlight informative spatial and channel relationships for improved discriminative capabilities. A separate global texture module, based on multi-head attention, captures broader structural cues that cannot be observed from local neighborhoods alone. The final classifier combines these local and global features to produce reliable predictions. Experiments conducted on a JPEG-compressed subset of the RAISE dataset, covering five types of manipulation, show that PRILNet consistently performs better than several existing forensic models, particularly under challenging low-quality conditions.

Few-shot node classification on graphs poses a significant challenge due to the limited availability of labeled data. Conventional solutions often rely on meta-learning frameworks to generalize from a few examples; however, these methods are typically computationally expensive, difficult to scale, and complex to implement. To address these limitations, we propose SAPNet, a simple yet effective two-stage framework that avoids meta-learning while maintaining high performance. In the first stage, Simple Graph Convolution (SGC) propagates node features across the graph to generate smooth and informative representations. The second stage trains a lightweight linear classifier on the few-shot support set and refines its predictions through Approximate Personalized PageRank (APPNP), which integrates neighborhood information to enhance label propagation. This decoupled “feature propagation followed by label propagation” design ensures both conceptual simplicity and computational efficiency. Experiments on benchmark datasets Cora, Citeseer, and Amazon-Clothing demonstrate that SAPNet consistently achieves state-of-the-art performance, outperforming standard GNNs and meta-learning methods, particularly under the challenging 1-shot setting. These results highlight SAPNet's robustness, scalability, and practicality for graph learning under limited supervision.

Digital image forgery is the process of manipulating an image to deceive or mislead observers with real or manipulated content. Median filtering is widely used to smooth images and obscure traces of tampering, making its detection critical for image forensics. However, identifying median filtering becomes more complex when additional operations, such as compression, resampling, or noise addition, are applied. To address this issue, we propose a lightweight convolutional neural network (CNN) model named SobelMNet, specifically designed for detecting median filtering in compressed images. The proposed model utilises a Sobel filter-based preprocessing step to enhance the residual differences between the original and manipulated images. These residuals, which capture subtle features indicative of median filtering, are analysed by CNN for classification. Further, the proposed model is evaluated on grayscale low-resolution images generated from the Dresden dataset for both binary and multiclass classification tasks. The model achieved a remarkable detection accuracy of 99.43% in median filter detection and outperformed state-of-the-art methods in various scenarios, including combinations of median filtering with Gaussian blur, resampling, and additive white Gaussian noise (AWGN) with an average accuracy of 98.32%. Finally, its lightweight architecture ensures computational efficiency, making it practical for real-world forensic applications.

Human Activity Recognition(HAR) is an essential field of research with numerous applications in human-computer interaction, security, surveillance, and healthcare. Even with significant improvements, recognizing activities in the real world, multiview scenarios remain a challenging task due to factors like scale variations, cluttered background, and variations in viewpoints. Therefore, in order to deal with these issues, this paper presents a robust and view-invariant framework for human activity recognition. This paper presents a framework to efficiently utilize both handcrafted and deep features for temporal modeling in a multiview human activity recognition. In order to make the feature as being scale and rotation in-variant we applied the Uniform Rotation-Invariant Local Binary Patterns (URI-LBP) which is a variant of local binary pattern. Subsequently, a streamelined VGG16 model with 10 layers is utilized to extract high-level spatial features while maintaining computational efficiency. These complementary features are then fed into a Spatio-Temporal Long Short-Term Memory (ST-LSTM) network, which captures temporal dependencies across video frames, enabling accurate recognition of dynamic activities. Comprehensive analyses on the benchmark datasets CASIA and IXMAS illustrate the efficiency of the proposed framework, with remarkable accuracies of 95.68% and 98.78%, respectively. The outcomes demonstrate the suggested methodology's resilience and versatility in difficult multiview human activity recognition situations.

Deepfake refers to an artificial intelligence-based technique to produce manipulated videos that look realistic. However, this good aspect of Deepfake sometimes pose serious threats to society as malicious intenet users exploit deepfakes to disseminate false information. That’s why a lot of research and applications are being developed in both the deepfake generation as well as detection field to mitigate its negative effect. Given the intensity of the work and its imporatnce in the field, there is a need to comprehensively review the existing literature and provide probable directions for future works considering the identified gaps and limitations. So, in this paper, a bibliometric analysis is conducted to provide a comprehensive analysis of this topic from various aspects such as influential authors who are active in this field and their collaboration, as well as countries and more specific institutions investing annually. In this study, the primary source of data is Web of Science (WoS), and we’ve employed keyword-based searches to retrieve pertinent information. The collected data from WoS has been scrutinized along several dimensions, including top document types, publication trends, source titles, and the productivity of various locations with an annual breakdown. To delve into collaborative efforts among institutions, authors, and regions, we utilized the VOSviewer tool. Additionally, the CiteSpace tool aided in identifying key focal points, research trajectories, and pinpointing significant shifts in citations for keywords. This comprehensive approach contributes to the in-depth analysis presented in this paper.

One of the most intriguing topics for researchers in the area of vehicle dynamics is the use of vehicle vibration to analyse dynamic behaviour. The primary goal of this research is to examine the dynamic behaviour of a road vehicle under various road conditions. In this study, an integrated computer model is developed through the bond graph methodology to analyse the dynamic behaviour of a large road vehicle. The initial stage is to create the 3D models of each vehicle base, suspension unit, tyre, seat, and joint as bond graph elements with power ports for physical connections. This work demonstrates the dynamic behaviour of the 4 DOF lumped human biomechanical model coupled with an 11 DOF full-car vehicle model using the bond graph simulation technique. Ride comfort conditions as per ISO 2631:1997 standards are used to validate the dynamic response of the bio-dynamic human coupled with a full car. Various other dynamic responses in the time domain have been captured in simulation for the conditions where the vehicle passes over different bumps of the same height and width of 1 m, 2 m and 3.7 m with a 40 km/h speed. Additionally, RMS responses have been studied for different vehicle velocities. The results presented in the paper show the fascinating behaviour of full-car vehicles.

In recent times, the evolution of Large Language Models (LLMs) has brought about transformative breakthroughs in many real-life applications including mental health. The continuous advancement of artificial intelligence and natural language processing techniques has led to noteworthy achievements, with LLMs showcasing considerable potential in the detection and prediction of various mental health issues. This review paper presents insights on the role of Large Language Models in mental health detection especially focusing on anxiety, depression, and stress detection. We first present a taxonomy for the categorization of current research based on several methods used, including prompt engineering, fine-tuning, and instruction fine-tuning. The core of this review focuses on the methodologies employed in recent studies where LLMs have been utilized for detecting mental health and analyzing the performance of various models on tasks like anxiety detection, depression detection, and stress detection. This study includes an analysis of different models, datasets, and algorithmic approaches, along with the integration of LLMs into healthcare systems, focusing on examining the strengths and limitations of different techniques highlighting the challenges, opportunities, and future gaps in mental health using large language models. Range of performance metrics including accuracy, precision, recall and F1-score have been employed to assess the models overall efficacy.

Anomaly detection involves finding data points that don’t follow the usual trends followed by the rest of the dataset. There exist several computationally efficient anomaly detection methods in data science. This paper explores and compares various machine learning and deep learning techniques. Anomaly detection can be a tricky task due to a multitude of reasons including highly imbalanced datasets, low amounts of data on actual frauds, unpredictable nature of the problem. The transaction in question does not necessarily need to be an outlier, which makes the problem contextual as well. Having performed fraud detection on credit card fraud and healthcare provider fraud datasets, we have been able to draw certain inferences as to which algorithms work best under which scenarios. Random Forest Classifier, XGBoost Classifier, Isolation Forests, Artificial Neural Networks, and Autoencoders are few of the prominent algorithms tested. After testing the algorithms, the results show that some algorithms work better with imbalanced datasets, while others can handle noise very well. We have also tried using certain dimensionality reduction methods in order to remove noise and improve efficiency, and certain sampling methods to handle class imbalances and then compared their results on these datasets.

The widespread application and development of technology have significantly improved people's lives. In recent years, deep learning has emerged as a powerful branch of artificial intelligence, enabling the creation of intelligent software and robots that assist with various daily tasks. Deep Learning, a subtype of Complex Artificial Neural Networks (ANNs) based on machine learning, is commonly used for accurate human behavior detection from smartphone data. In this paper, we propose a model to identify human activity using ConvLSTM, DenseNet, and LRCN. The statistical model is built on a substantial dataset of videos sourced from publicly accessible datasets UCF50 and HMDB51, which cover a wide range of activity variables. Our results demonstrate that the DenseNet-based model outperforms other models in terms of accuracy.

In recent times, the volume of video surveillance data has been rapidly increasing, and processing is required to detect various kinds of activities in these data. Most human activity recognition methods primarily focus on detecting normal human activities; however, identifying and detecting suspicious activities is equally important. We propose a two-stream framework for detecting abnormal human activities in real-time videos. The proposed method uses the first stream to extract features related to spatial characteristics and the other stream to extract the temporal features from the videos. The spatial features focus on the contextual details in the frame, including color, texture, and other robust features. The temporal features focus on the features related to the changes in frames with respect to the time, which play a vital role in the detection of activity. The proposed framework leverages the Xception model to extract the features in the spatial and temporal streams. The extracted features are integrated for further processing and detection. To process the integrated features and detect the activity, we train a multi-layer bidirectional long short term memory network using the extracted features. The proposed framework is trained in an end-to-end manner to accurately estimate human activity by learning patterns and forming correlations. The effectiveness of the proposed approach is evaluated by assessing the popular UCF Crime dataset, which encompasses a variety of criminal activities, and using performance measures such as recall, precision, and accuracy. Experimental results indicate that the proposed approach outperforms other contemporary methods, achieving an accuracy of 86%. (c) 2023 SPIE and IS&T

Every day, the volume of textual information is growing at an accelerated rate. A portion of this data is presented as product, movie, game, and other review content. These reviews fall into two general categories: favourable and unfavourable. Since there are a lot of reviews for popular products, a classifier that can categorise reviewers as positive or negative is necessary. This work compares the performance of BERT with other widely used approaches for review classification. Specifically, we combine BERT with several deep learning algorithms (BiGRU, 1-D CNN, and GRU-CNN). Using a pre-trained BERT model, we first turn the content into numerical vectors. Next, we feed the embeddings into several models and evaluate how well they perform. We evaluate the performance based on f1 macro, accuracy, precision, and recall and find that the BERT+CNN_GRU model outperforms the widely-used models for review classification.