Cloud Computing has fundamentally transformed how organizations worldwide provision and manage infrastructure, while at the same time introducing notable challenges for digital forensics and incident response. The traditional forensic approach relies heavily on physical access and static storage. This approach does not work well with cloud environments, as everything is virtual and dynamic, and evidence is distributed across many services and is volatile in nature. This paper presents a flexible cloud-focused forensic analysis framework. The framework integrates data from various AWS services, including CloudTrail, VPC Flow Logs, AWS Config, IAM Activity, and Security Configurations. It also gathers data from virtual machines, such as system logs and live memory snapshots, all managed by an API-driven framework. This framework helps investigators connect different types of evidence and build clear timelines, making it easier to identify security problems with less manual work. We tested this framework in a typical AWS environment. The results indicate that it can collect and link evidence from different parts of a system and assist in reconstructing events during suspicious activities. This work suggests that the framework can improve forensic processes and points to future research in automated forensics for both single-cloud and multi-cloud environments.
This paper gives a complete evaluation of Rhythm-Net, an innovative deep learning (DL) technique for the diagnosis of heart disease from multimodal information (ECG, PPG, and clinical information). The model’s performance is tested using standard metrics like accuracy, precision, recall, F1-score, ROC-AUC, and blockchain-based metrics like latency and throughput. Comparative evaluation is performed against both conventional machine learning (ML) models (Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), K-nearest neighbor (KNN)), state-of-the-art DL (benchmark) models (Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (BiLSTM) and advanced existing models (DeepHeart, HeartNet, MultiResCNN, ResNet-1D) by comprehensive tabular and graphical results. An ablation study also examines the effect of multimodal fusion, privacy-preserving blockchain incorporation, and important architectural elements such as attention and residual connections. The results show the better diagnostic performance of Rhythm-Net, with high robustness and privacy preservation, and a balanced presentation of advantages, limitations, and clinical readiness.
Text-to-sign language translation remains a challenging task due to structural differences between spoken or written languages and sign languages. Recent research has shifted from rule-based approaches to neural network architectures. It has significantly improved text to gloss translation. This study has four main objectives. The first objective is to conduct a comprehensive literature review to identify widely used neural models for this task and to define the proposed model. The second objective is to design an algorithmic framework for the proposed model. The third objective is to evaluate the proposed model against baseline models and state-of-the-art methods. These experiments are conducted on the benchmark ASLG-PC12 dataset. The results of Research Study 1 demonstrate that the proposed GlossTransformer model consistently outperforms baseline models. It achieves a training accuracy of 99.45% and a testing accuracy of 99.20% where 80:20 split is used for training and testing split. An 80:10:10 train–validation–test split is used in Research Study 2. The results of Research Study 2 are assessed using token-level precision, recall, F1 score and sequence-generation metrics show that the proposed model achieves competitive performance. On the test split, GlossTransformer reports 99.39% accuracy, 96.83% precision, 97.50% recall, 97.09% F1 score, BLEU-1 97.49, BLEU-2 96.55, BLEU-3 95.62, BLEU-4 94.70, WER 2.55, and ROUGE-L 97.51 and perform comparably to existing state-of-the-art methods. Future work is needed to focus on broader testing with other datasets and improved generalization analysis. This framework can later include sign language output and speech recognition to better support people with hearing disabilities.
Speaker diarization determines who talked when by segmenting a recording according to speaker identity and labeling each segment uniformly throughout the entire broadcast. The task has undergone a nearly complete transformation since 2019. The earlier pipeline methods were unable to resolve speech in which voices overlap because they performed voice activity detection, feature extraction, and clustering separately. End-to-end neural diarization (EEND), on the other hand, reframes the task as a frame-level permutation-invariant multi-label classification problem. Instead of treating speaker assignment, overlap detection, and segmentation as distinct subtasks, this unified approach allows a single model to jointly learn each of these tasks. Instead of providing a comprehensive survey, this review, which covers the years 2019–2026, provides an in-depth perspective. In addition to examining recent integrations with foundation models, such as DiariZen, Sortformer, EEND-M2F, and SpeakerLM, it tracks the evolution from typical clustering-based pipelines to self-attention-based EEND architectures, encoder–decoder attractor models, multi-scale decoders, and streaming approaches. The lowest reported diarization error rate (DER) has dropped from roughly 26.2% (VBx, re-evaluated in 2024) to 15.1% (DiariZen, 2025) when assessed under a consistent no-collar approach on the AMI single-distant-microphone (SDM) corpus. In order to provide a harmonized comparison of DER values brought onto a standard evaluation footing under a stated approach, we assess reported performance across eight benchmark datasets. We expand the analysis to language diarization (LD) in code-switched and multilingual contexts, finding fundamental issues with SD that are addressed independently in previous surveys. Overlapping speech at three or more simultaneous speakers, scalable speaker counting, real-time deployment, annotation dependency, and domain generalization are among the open issues that we systematically catalog and relate to existing solutions and gaps. We conclude by outlining intriguing avenues for future research, such as multimodal integration, joint diarization and transcription, parameter-efficient foundation model adaptation, and fairness-aware evaluation. According to our analysis, the field is at a turning point where it can perform almost as well as humans in controlled environments but remains limited to acoustic complexity, linguistic diversity, and deployment realities.
Veterinary diagnostics is a field that demands AI-based solutions suitable for resource constraint environment. Traditional animal disease detection methods have significant scope for AI adoption to automate certain aspects of the diagnosis process. In this study, twelve CNN-ML hybrid models are developed, leveraging the power of Convolutional Neural Networks (CNN) combined with traditional machine learning approaches. Pre-trained models are employed for feature extraction paired with traditional ML classifiers such as Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boost (XGBoost) and AdaBoost. These twelve hybrid models were categorized based on their CNNbackbone: category I (DenseNet121+ML models), category II (VGG16+ML models) and category III(InceptionV3+ML models). The top-performing hybrid models from each category i.e., DenseNet121+RF (96.00% accuracy) from category I, VGG16+SVM (93.5% accuracy) from category II and InceptionV3+SVM (92.60% accuracy) from category III are selected for ensemble modeling. The weighted ensemble achieved an accuracy of 95.5%, while the voting ensemble is able to achieve the accuracy to 96.62%. The ensemble model is shown to be more generalized and robust compared to the individual hybrid models. Additionally, as training time was a crucial consideration in resource-constrained environment, we concluded that weighted ensemble model is efficient as its training time is only 0.97 seconds.
PCOS is a type of endocrine dysfunction that impacts women's reproductive systems. This illness reduces the likelihood of becoming pregnant. The primary cause of this disease is the female body's suspiciously increased levels of the male hormone known as androgen. The normally occurring hormones in females are impacted by the aberrant function of this androgen. Atypical menstrual cycles, excessive menstrual blood, acne, and excessive weight are among the signs of PCOS. In certain cases, these conditions can worsen and result in major health issues like type 2 diabetes and ovarian cancer. Early PCOS detection may help manage the condition. In light of this, this work presents a hybrid model with an enhanced resampling technique for PCOS detection that consists of five steps. These include feature extraction and detection, data augmentation, data sampling, and preprocessing. Data normalisation is done as part of the preparation stage of data, which improves dataset quality for more insightful analysis. Afterwards, a better SMOTE framework is suggested for dataset sampling. The min-max data augmentation procedure is used to enhance the sampled data. Different features are then extracted, including those based on mutual information, statistics, and entropy. Lastly, a hybrid classification model that combines the deep maxout classification model and BiGRU is employed for the classification process. Lastly, several measures are used to assess the performance of the suggested task.
This paper introduces a blockchain-based electronic voting system that aims to improve the security, transparency, and efficiency of the voting process. Conventional voting systems are susceptible to security attacks, fraud, and lack of transparency, resulting in public distrust of the democratic voting process. Through the use of Ethereum blockchain technology and self-executing smart contracts, the system proposed here automates essential voting processes like voter registration, candidate management, and vote validation, guaranteeing decentralized, tamper-proof, and immutable operations. Two implementations are considered—one that is entirely based on blockchain for a totally decentralized election process and another combining blockchain with a local database for enhanced accessibility and efficiency. The system utilizes cryptographic hashing and digital signatures to safeguard voter authenticity, while the immutability of blockchain prevents vote alteration. The decentralized approach mitigates risks such as double voting and unauthorized access, ensuring voter verifiability and auditability of election results. It also facilitates voter verifiability, decreases operational expenses, and promotes confidence in the democratic process. Experimental deployment and testing, using Solidity for smart contract development and the Truffle framework for blockchain interactions, validate the system’s robustness. Gas cost analysis highlights the computational efficiency of various operations, demonstrating feasibility for large-scale elections. This research showcases blockchain's potential to transform e-voting by guaranteeing fair elections, providing a scalable and secure alternative to traditional electoral systems, reducing operational costs, boosting voter turnout, and minimizing fraud risks, opening the door to future developments in digital democracy.
Cloud security has gained widespread popularity across various industries due to a variety of services that offer improved functionality, scalability, and affordability. However, the security of cloud environments has become increasingly critical due to the increasing number of attacks. Traditional methods of security monitoring often rely on predefined rules and thresholds which may not adapt well to evolving threats or complex attack patterns. An approach for detecting anomalies in the cloud by analyzing patterns in AWS (Amazon Web Service) CloudTrail logs has been presented. Although there are other cloud service providers available, AWS is chosen for anomaly detection due to its popularity. Analyzing such a vast number of logs without automation is complex, and detecting anomalies within this large dataset is a crucial task. Artificial intelligence (AI) is chosen for this task because it can spot patterns and anomalies that are too difficult to find manually, making detection more accurate and efficient. This paper examines the various supervised and unsupervised machine learning for accurate anomaly detection. By automating this process, the system aims to provide real-time alerts, enabling prompt responses to potential issues. The performance of the system is evaluated using metrics like precision, recall, and accuracy score. Our results demonstrate 97% detection accuracy. The results indicate that the proposed solution can effectively identify anomalies with high precision, demonstrating its potential to improve cloud security and operational management.
Human Centered Computing is a rising field of research having goal of understanding behavior of human being. In this field, Sign Language Recognition is noteworthy application with many challenges. It bridges the communication gap that separates hearing individuals from deaf people. In this article, a computer vision techniques based application for recognising Indian Sign Language (ISL) is presented. Convolutional Neural Networks (CNNs), the most remarkable kind of algorithms for deep learning, are used in its construction. Initially, the self built dataset containing training images has been used to train the system. The system interprets a variety of live camera gestures into text and voice. To further enhance the system's usability, a graphical user interface (GUI) is also built. Additionally, it also offers rotation, translation, and scale invariance. The CNN's weight sharing and sparse connectivity make the suggested system incredibly effective in the context of storage and performance. It is found to be an extremely accurate system with a 99.44% recognition rate during the testing.
Thyroid nodules are one of the most common abnormalities in the thyroid gland, which are often harmless in nature (benign), but in a few unfortunate instances, they may be fatal (malignant). This review explores recent advancements in artificial intelligence (AI) applied to thyroid cancer detection and classification, with a focus on machine learning, deep learning, and image processing techniques. We provide a comprehensive evaluation of AI applications across key imaging modalities—Ultrasonography (USG), Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Single-Photon Emission Computed Tomography (SPECT) and Positron Emission Tomography (PET/CT)—as well as cytopathological analysis using Fine Needle Aspiration Biopsy (FNAB). By critically examining studies on AI-driven preoperative assessments, we highlight improvements in diagnostic accuracy, sensitivity, specificity and efficiency. This review also identifies current limitations in AI applications, including technical challenges and unresolved issues that hinder widespread clinical adoption. Although significant progress has been achieved, the integration of AI in clinical settings remains limited, as AI-based outputs currently serve as supportive tools rather than definitive diagnostic evidence. We discuss the potential of AI to transform thyroid cancer diagnostics by enhancing reliability and accessibility, while addressing the need for further research to develop a unified, robust and clinically trustworthy AI framework for thyroid cancer diagnosis.
The patterns of providing health-centric services have transformed extremely with the enhancement along with innovations in mobile and wireless communication technologies subsuming the Internet of Things (IoT). Due to the rapidly increasing attack, the doctors were not provided with an accurate alerting mechanism by the prevailing health monitoring system. Thus, by utilizing the Exponential Linear activation Units-centred Bidirectional Long Short Term Memory (ELUS-BiLSTM) technique, a novel healthcare monitoring along with an attack detection system is proposed in this work. Attack detection, Data security, and Patient health monitoring are the three primary phases incorporated in the proposed methodology. Initially, from the patient, the data are collected, and then the features are extracted in the attack detection phase. Next, the features being extracted are inputted to the ELUS-BiLSTM classifier where the data is classified as attacked or non-attacked data. After that, by utilizing Skew Tent Elliptic Curve Advanced Encryption Standard (STECAES), the non-attacked data is encrypted whereas the attacked data is stored in the log file. Lastly, to generate the fuzzy rules, the encrypted data is utilized; subsequently, the alert message is sent to the doctor. The experiential outcomes displayed that when analogized with the prevailing methodologies, the proposed model obtained better outcomes.
As digital technology for illness diagnosis and analysis has advanced, medical images are sent over the Internet. Cloud computing plays a major role for low-cost data storage and data sharing. In the healthcare industry, data security and privacy are key issues with cloud computing. It is imperative that healthcare professionals make sure that patient data is safe against hackers, unauthorized access, and thrift. To secure the confidential data and store the huge amount of data, encryption and compression techniques are used. This paper provides an enhanced optimized encryption and efficient hybrid compression strategies based on cloud environment. The proposed model involves various operations such as generate optimal key, encryption, compression, decompression and decryption. In order to transfer the data using high speed cloud data retrieval, we initially propose the Huffman Fano Hybrid Entropy approach. In the next step, the proposed model performs Elliptic Curve Coding based encryption technique to secure compressed medical image transmission. Here the shared secret keys are generated optimally by dynamic group based cooperative optimization algorithm, which makes use of the encryption quality measures and is called DGBCO–ECC model. On the receiving end, the image is decrypted and decompressed. The proposed model performance is validated by exploiting various parameters namely Mean Square Error, Standard Deviation of the Mean Error, Universal image quality index, Structural Similarity Index, Entropy, Peak Signal to Noise Ratio, Compression Ratio, Data Rate Saving and compression time. When the experimental outcome is contrasted with existing methods, it is found to perform better.
Worldwide, the most enormously produced fruits, including bananas, papayas, mangoes, and guavas, are found in India. Over the years, agricultural production in India has consistently increased. There is still a massive gap between per capita demand and supply due to losses, including post-harvest. With adequate processing facilities, a clear scope exists to reduce this post-harvest wastage. In recent years, research in cutting-edge technology like computer vision (CV), Artificial Intelligence, and image processing has played an important role in sorting as well as grading fruits. Fruits in similar colors and textures increase the difficulty of identification. Deep learning networks are used to adapt and recognize complex patterns, especially in visual tasks. Utilizing deep transfer learning facilitates achieving excellent results expeditiously. This paper uses the deep transfer learning approach to classify fruits with similar color and texture, namely guava, avocado, lime, apple, pear, mango, and pomelo sweetie. This study introduces a novel model derived from integrating DenseNet, MobileNet, and EfficientNet architectures. The model’s performance is systematically assessed using different optimizers, contributing to a comprehensive evaluation of its efficacy. Simulation findings indicate that MobileNetV1 when paired with the Adam optimizer, surpasses other models in terms of training time, accuracy, and testing time.
In low- and middle-income countries, oral cancer is becoming more common. One factor delaying the discovery of oral cancer in rural areas is a lack of resources. To stop the disease from spreading, it is essential to quickly obtain information about any cancers. Therefore, it is essential to carry out early identification before it spreads. Primary screening is maintained in this study. Furthermore, deep neural network-based automated methods were used to produce complex patterns to address the challenging issue of assessing oral cancer infection. The goal of this work is to develop an Android application that uses a deep neural network to categorize oral photos into four groups: erythroplakia, leukoplakia, ulcer, and normal mouth. Convolutional neural networks and K-fold validation processes are used in this study’s methodology to create a customized Deep Oral Augmented Model (DOAM). Data augmentation techniques including shearing, scaling, rotation, and flipping are used to pre-process the images. A convolutional neural network is then used to extract features from the images Optimal configurations of max pooling layers, dropout, and activation functions have resulted in the attainment of maximum accuracies. By using the ”ELU” activation function in conjunction with RMSProp as the optimizer, the model achieves 96% validation accuracy, 96% precision, 96% F1 score, and 68% testing accuracy. The model is then deployed in TensorFlow Lite using an Android application.
The detection of multiple tiny objects from diverse perspectives using unmanned aerial vehicles (UAVs) and onboard edge devices presents a significant challenge in computer vision. To address that, this study proposes AeroNet, a lightweight and efficient detection algorithm based on YOLOv7 (You Only Look Once version7).This algorithm features the LHGNet (Lightweight High-Performance GhostNet) backbone, an advanced feature extraction network that integrates depth-wise separable convolution and channel shuffle modules.These modules enable deeper exploration of network features, promoting the fusion of local detail information and channel characteristics. Additionally, this research introduces the LGS(Lightweight Gradient-Sensitive) bottleneck and LGSCSP(Lightweight Gradient-Sensitive Cross Stage Partial Network) fusion module in the neck to reduce computational complexity while maintaining accuracy. Structural modifications and adjusted feature map sizes further enhance detection accuracy. Evaluated on the SkyFusion dataset,this method demonstrated a 25.0% reduction in parameter count and a 12.8% increase in mAP (0.5) compared to YOLOv7. These results underscore the effectiveness of this proposed approach in improving detection accuracy and model efficiency through the proposed enhancements.
Medical Visual Question Answering (MedVQA) is a dual research field that intersects medical imaging and natural language processing, for better interpretability and accessibility of medical image data.Medical image quality is paramount for accurate diagnostics and subsequent medical visual question answering (MedVQA) tasks. This research focuses on applying and then analyzing results of different denoising methods on VQA-RAD, Medical VQA dataset to enhance quality of images. This study explores effectiveness of different traditional and deep learning based methods to reduce noise within medical images, thereby improving the accuracy and reliability of MedVQA task. We applied different traditional denoising filtering methods such as, gaussian filter, median filter, average filter, bilateral filter and convolutional autoencoder (CAE) based on deep learning on a VQA-RAD dataset to compare effectiveness of each denoising methods to improve image quality. Through comprehensive experiments and evaluations, this paper demonstrates that the convolutional autoencoder is potentially enhancing quality of medical images with an emphasis on preserving essential diagnostic information while suppressing unwanted noise with compare to other traditional denoising filters. The denoised images are then employed as input to improve accuracy for MedVQA tasks. The results of this research will help in optimizing medical imaging pipelines, ultimately benefiting clinical decision-making and healthcare outcomes.
Developing self learning model for a game is challenging as the environment keeps changing all the time and therefore require highly intelligent models which can make decisions depending on the environment in real time. The agent has to learn the environment and takes action based on the inference. Based on the action, a positive or negative reward is given to the agent. The agent again learns from the reward and enhances / trains itself to behave better in the environment. This work aims to train an agent using deep reinforcement learning algorithms to play a multiplayer online game like SLITHER.IO. We use an OpenAI Universe environment to collect raw image inputs from sample gaming as training data. Agent learns the current state of the environment and the position of the other players (snakes). Then it takes action in the form of direction of its movement. To compare our model to other existing systems and random policy, we propose to use deep Q-learning and other actor critic approaches such as Proximal Policy Optimisation (PPO) with reward shaping and replay buffer. Out of all these algorithms the PPO agent shows significant improvement in the score over a range of episodes. PPO agent learns quickly and its reward progression is higher when compared to other techniques.
The Electronic Health Record (EHR) is a digitalized solution to support the health care facility, irrespective of levels and sizes to improve patient care system by eliminating the paper based medical records.Standardization of EHR improves the easy sharing of health information between various levels of health care system. The availability of the patient’s data in a timely fashion can contribute to the improvement of patient’s information and performance of the Health Information System. Current health care information systems of the hospitals are usually isolated from each other as most of the hospitals and health care institutions have their own format to create EMR (Electronic Medical Records) to serve the purpose of treating the patient.Standard coding makes it simple to share health information, lowers uncertainty, enhances workflow, and makes it easier to accurately analyze data related to health care.During patient registration or hospital visit, ID proof like Aadhar Number isused as a universal patient identifier. Healthcare user authentication is archived at database level through valid user name and password.The cloud server checks the credentials against a user store of the database for validation as illustrated in Algorithm-1.The primary function of the attribute based access control (ABAC) provided by Algorithm-2 is to authorize access for healthcare users. The hospital authorities obtain the patient's agreement in the first stage, and the loop is continued by using the value YES.The role based access control (RBAC) given in Tables-II and III is one of the best method for highly complex and huge management system. All this process standerdize EHR and its controlled access safe and secure.
User authentication is an important process that ensures only authorized individuals can access a system or network infrastructure. This process protects users’ sensitive information from unauthorized access and prevents any unwanted tampering. In this research paper, a unique method for authenticating users based on typing dynamics has been proposed, aimed to enhance the security of systems and networks by verifying user identities. This method involves calculating the period of time each key is pressed and released. The study includes data collection, feature extraction, model training, and performance evaluation by measuring the accuracy and precision of the trained model. We evaluated three deep-learning models to test the proposed method and determine its accuracy, precision, and superiority among the three models. Based on the findings of this research, we are able to present an algorithm which outperforms the other two considered for the experiment. Also, a comparative study is been presented after the first evaluation which involves assessing the accuracies for different lengths of the password. Additionally, charts and graphs were carefully employed to ensure precise representation and effective visualization of the data.
In India, one of the primary agricultural practices is the production of bananas. A prevalent issue in farming is that the crop has been impacted by multiple illnesses. Disease identification in bananas has been shown to be more difficult in the field because the fruit is prone to various diseases and causes farmers to suffer significant losses. Consequently, this study aimed at developing an automatic system for the early detection and classification of banana plant diseases using deep learning. Three pre-trained convolutional neural network models MobileNet, VGG16, and InceptionV3 are used to classify banana disease images. The banana disease images dataset from the PSFD-Musa Dataset is utilized for training, validation, and testing. The proposed system is developed and checked to classify banana plant disease photographs into one of seven categories. The MobileNet achieved an accuracy of 96.72%, VGG16 an accuracy of 55.68%, and InceptionV3 an accuracy of 63.65%.
