Mathiyalagan Siva Ramkumar

Weld component durability is essential in industrial production processes, where unidentified defects can cause serious operational failures. Automated weld flaw identification through deep learning technologies provides a dependable solution that eliminates the need for human inspection. This research introduces a new deep learning framework that integrates AlexNet with Gated Recurrent Unit (GRU) Network and Greylag Goose Optimization (GGO) method to achieve effective hyperparameter tuning. The proposed framework tests its performance through the RIAWELC dataset which contains 24,407 weld radiograph images divided into four categories: porosity, crack, lack of penetration, and no defect. The weld images receive quality improvement through advanced preprocessing techniques, which use median filtering, Contrast Limited Adaptive Histogram Equalization, and data augmentation methods. AlexNet extracts discriminative spatial features from visual data while GRU learns the contextual relationships among these features. The GGO technique enables the suggested framework to achieve its optimal performance through the GGO method which optimizes critical hyperparameters such as learning rate, dropout, batch size, kernel size, and GRU hidden units. The experimental results demonstrate that the suggested framework outperforms all existing methods by achieving 99.3% accuracy in binary classification and 98.9% average accuracy in multi-class classification.

Bone cancer is an aggressive and rare disease but can often comprise osteosarcoma, chondrosarcoma, and Ewing sarcoma, and for their cure, early diagnosis is essential cancer diagnosis needs very accurate classification methods because of the complexity and variability in structure of tumors. The current research proposes a new Dilated Convolutional Network with Pied Kingfisher Optimization (DCN-PKO) for better detection of bone cancer. The suggested framework combines Potential Low-rank Decomposition and Anisotropic Guided Filtering (PLD-AGF) as pre-processing to improve image clarity and maintain essential structural details. Modified Inception MV4 (MI-MV4) is used for feature extraction, which enhances pattern identification in medical images. Additionally, Parrot Optimization (PO) is utilized for feature selection to minimize computational complexity while maintaining vital features. The classification process employs Dilated Convolutional Networks (DCN) for strong feature representation, after which Pied Kingfisher Optimization (PKO) is employed to optimize weight parameters, allowing high accuracy with little error. The model is tested on the TCGA bone cancer data set, containing MRI and CT scans of osteosarcoma, chondrosarcoma, and Ewing sarcoma among other bone tumors. Comparative study using advanced techniques proves that DCN-PKO performs better and reaches 98.7% accuracy, 98.6% precision, and 98.8% recall at an error rate of just 0.2%. These findings verify that the proposed method greatly enhances classification accuracy, decreases computational expense, and promotes early diagnosis, which makes it a promising approach for real-world medical applications in bone cancer diagnosis.

Medical devices, software, and networks have joined the Internet of Medical Things (IoMT) and have enhanced healthcare services but pose threats. IoMT devices are vulnerable to cyberattacks that can affect patient data as well as the operations of the healthcare organization, which requires an enhanced and effective intrusion detection system (IDS). This paper introduces a novel approach for detecting cyberattacks in IoMT environments using a deep learning model named pyramid attention network with the Green Anaconda algorithm (PANet-GAO). The approach has three steps as follows: 1) MinMax scaler normalization to process the data, 2) multiple discrete orthonormal S-transforms (MDOSTs) to extract the critical temporal-frequency features, and 3) pyramid attention network with the Green Anaconda algorithm (PANet-GAO) for accurate cyberattack detection. The MDOSTs can extract the dynamic information from the network traffic data in the time-frequency domain, and the feature space of the PANet model is optimized by GAO, which minimizes the feature redundancy to improve the classification accuracy. The proposed model was tested using the WUSTL-EHMS dataset, which mimics real-life IoMT traffic and different cyberattacks. The proposed PANet-GAO model's precision was 99.76%, its accuracy was 99.86%, its F1-score was 99.78%, and was better than traditional machine learning algorithms. Therefore, this work establishes that the proposed PANet-GAO approach can effectively protect IoMT environments from emerging cyber threats in real-time.

The rapid expansion in Internet of Things (IoT) devices has fueled innovation in automation and real-time monitoring but in parallel increased exposure to sophisticated cyber-attacks. Traditional detection methods are not able to handle the complex, noisy, and hierarchical nature of IoT communication and therefore limited to classification. This research proposes an IoT Cyber Defense mechanism: Node-Centric Hyperbolic Capsule Graph Neural Network Approach to Threat Detection and Classification (NCGNN-HAGN). First, IoT traffic data from the NSL-KDD and BoT-IoT datasets are integrated to represent a wide range of attack patterns. Second, Multiple Local Particle Filtering (MLPF) methods are employed to remove noise and improve the quality of communication. Third, the Second-order Synchro-Extracting Wavelet Transform (SSWT) is used for extracting timefrequency and spectral features that characterize burst patterns and frequency anomalies in IoT traffic. Artificial Jelly Optimization (AJO) chooses the appropriate features, decreasing dimension and increasing classification efficiency. The optimized features are fed into Node-Centric Capsule Graph Neural (NCGNN) with Hyperbolic Attention Graph Network (HAGN) integration, extracting hierarchical and multi-relational relationships between IoT nodes. Chef Leader-Based Optimization (CLBO) optimizes NCGNN parameters, advancing learning accuracy. Results show that this method attains 99.9% greater accuracy, 99.8% greater precision, and 5.87% less computation time than current methods, proving its effectiveness, scalability, and robustness in IoT cyber-attack detection.

Massive multiple-input multiple-output (maMIMO) networks have gained significant attention in recent years with their capability to enable efficient communication through offering base stations (BS) access to an enormous number of antennas to serve multiple user equipment (UE) at a time. One of the key challenges in maMIMO networks is optimal power allocation among UEs keeping the total power constraint at the BS intact. Despite the fact that deep learning-based techniques offer rapid power allocation, they are vulnerable to adversarial attacks on input data, leading to incorrect resource allocation. To address the above problems, this current study proposes a model named Resilient Power Allocation against Universal Attacks in Massive MIMO using Multi-Stream and Component-Aware Graph Networks (PADOM-AKGNTO). To start with, input data in terms of user equipment (UE) positions along with their corresponding optimal power allocations is collected and pre-processed via an Adaptive Fast Desensitized Kalman Filter (AFDKF) for the purpose of filtering noise and uncertainty reduction. For predicting optimal power allocation, a Multi-Stream Fusion Spatiotemporal Graph Convolutional Network (MSF-STGCN) is used to determine spatial and temporal relations among UEs in a dynamic scenario. For providing robust defense against adversarial attacks capable of manipulating UE position data, a Multi-Component Attention Graph Convolutional Neural Network (MCAGCN) is utilized as a resilient detection and mitigation module. A Termite Life Cycle Optimizer (TLCO) is also utilized to optimize weights and parameters of the models MSF-STGCN and MCAGCN for the purpose of improving accuracy and reducing convergence speed. The tested and validated PADOM- AKGNTO model employs different performance metrics and attains a prediction accuracy of 99.99%, defense success rate of 98.50%, and lowers the attack success rate significantly to 5.00%. The outcomes confirm that the suggested method facilitates efficient, accurate, and stable power distribution in hostile environments, thereby ensuring secure and reliable communication in maMIMO systems.

The rise of cyber-attacks in smart Internet of Things (IoT) settings has necessitated scalable and efficient attack detection mechanisms. This work introduces SelfEvolving Middleware for Cyber Threat Detection in Smart IoT Ecosystems Using Cycle-GANs and Bio-Inspired Optimization (Cycle-GANs-TLCO). Network traffic traces are first fetched from the IoT-23 and Edge-IIoT datasets. Preprocessing through Confidence Partitioning Sampling Filtering (CPSF) is accomplished to remove noise, balance attacks, and scale the dataset to improve learning. The Gooseneck Barnacle Optimization Algorithm (GBOA) is utilized for the selection of discriminative network traffic features, while being computationally light. For the detection of cyber-attacks, the Unsupervised Cycle-Consistent Generative Adversarial Network (UCCGAN) is implemented with adversarial learning for pinpointing subtle attacks. Hyperparameters are further tuned with Termite Life Cycle Optimization (TLCO) to better improve model accuracy and stability. Experimental evaluation evinces an increment in attack detection accuracy by 17.85%, 23.41%, and 29.74% against state-of-the-art competitors. The deployment of the middleware showcases an enhancement of CPU use by only 2.1%, with memory usage increasing to 0.42 GB, to maintain resource efficiency. Power consumption is minimized by 15.9% and 11.2% on NVIDIA Jetson and Raspberry Pi, respectively, to reinforce the feasibility for real-time use in IoT security applications.