Phishing emails are a serious cybersecurity risk because they take advantage of people's weaknesses to obtain private information. Both conventional heuristics and classical machine learning techniques are undermined by their intrinsically noisy, obfuscated, and constantly changing character. Because they are unable to learn invariant, discriminative representations and deliberately suppress noise, even the most advanced deep learning models have limited robustness. In this paper, we provide a novel noise-augmented, attention-guided contrastive learning framework for reliable phishing email detection in order to overcome these difficulties. However, we introduce two crucial novelties: (i) the noise augmentation policy, producing perturbed views of an email to account for variability and obfuscation in the wild, and (ii) an attention-guided contrastive learning method to focus on more informative features and discard noisy ones in contrastive representation learning. By leveraging this synergy, the model can obtain class-discriminative and noise-invariant embeddings, guaranteeing accurate identification even in the face of extreme adversarial perturbations. Comprehensive evaluations on the publicly accessible Phishing Email Detection dataset show that the proposed framework routinely achieves accuracy gains of at least 1.5% and 1.22%, respectively, over twelve cutting-edge baselines. These findings confirm the efficacy and superiority of the suggested method for acquiring reliable, noise-resistant representations for precise phishing detection in practical settings.

This paper proposes a novel hybrid approach called Multiplex k-shell and Vertex Cover Integrated Independent Set (MKVI) for influence maximization in a multiplex network. Here, the concepts of k-shell decomposition, vertex cover, and independent set are extended to multiplex networks and combined to select influential nodes (seed sets). The average influence spread of the selected seed set is computed by simulating the independent cascade model with three different influence probabilities (p), where the p values are set to 0.01, 0.05, and 0.1. The proposed MKVI reports a dominant average influence spread in comparison to six state-of-the-art approaches: Degree centrality (DC), Betweenness centrality (BC), Eigenvector centrality (EC), K-shell coefficient (KC), Cost-Effective Lazy Forward selection++ (CELF++), and Reverse random walk (RRW), across six datasets: CElegans, Drosophila, CKM-Physicians, CS-Aarhus, EUAir_Multiplex, and Padgett-Florence-Families_Multiplex_Social. MKVI also reports a closer average influence spread to that of multiplex networks using the learning automata (MISM-LA) method while requiring a much smaller seed set size.

Detecting code smells is very important for increasing software maintainability and lowering the technical debt of large-scale software systems. Traditional machine learning methods rely heavily on manually engineered features and, as a result, can struggle to generalize across projects due to domain differences and class imbalance in the datasets. However, although transformer-based pre-trained models have shown great promise in understanding the semantics of source code, there has been limited investigation into how well they perform across different datasets, particularly balanced versus imbalanced ones. In this study, we compare the performance of baseline machine learning models and transformer-based models for detecting multiple types of code smells on two heterogeneous datasets with different distribution properties. From the analysis, we see that the degree of imbalance in the datasets and the differences between the two domains significantly affect the performance and generalization of the various models. Our experimental results show that whilst transformer-based models outperform baseline machine learning models, the extent of their advantage varies with dataset characteristics; therefore, transformer-based models do not generalize well across projects. We have also found that providing domain-specific fine-tuning strategies can improve adaptability and detection performance in real-world use. This study provides insights into dataset characteristics, model behavior across domains, and the need for adaptive learning approaches to develop robust, generalized code smell detection systems.

Smart manufacturing systems based on software become very dependent on embedded control applications, industrial automation software and real-time decision mechanisms to ensure the reliability of production. Nonetheless, software failures that may arise in such systems are usually not foreseeable because datasets are highly unbalanced, in which the number of faulty software modules is much less than the number of non-faulty modules. This paper will present an adaptive hybrid learning paradigm of software fault prediction in smart manufacturing systems through class imbalance optimization. The proposed framework is composed of three significant modules: (1) Feature Extraction Module that is used to extract software metrics, execution logs and operational indicators; (2) Class Imbalance Optimization Module that applies Synthetic Minority Oversampling Technique (SMOTE) and Borderline-SMOTE to equalize minority fault classes; and (3) Hybrid Prediction Module that implements the use of the random forest, support vector machine, Multi-Layer Perceptron and Bayesian Network through weighted voting classification. NASA software defect repositories and PROMISE Repository datasets are benchmark software defect datasets on which models are validated. As shown by the results of the experiment, the proposed framework has a prediction accuracy of 97.1%, precision of 96.2 and a recall of 95.8, which is superior to traditional single classifiers. The framework enhances the detection of minority faults, minimizes false negatives and facilitates predictor software maintenance in industrial automation conditions. Future directions of work are explainable fault prediction and real-time deployment in edge-based manufacturing systems.

Reliable human recognition still remains a challenging task from a security perspective. Changes in illumination limit a Unimodal Biometric Identification system that uses a single biometric modality, pose variations, occlusions, and behavioral variations. To overcome the limitations mentioned above, we propose a multi-biometric approach combining face and gait recognition. In this research, we propose an adaptive multibiometric identification system that combines face and gait information via a hybrid deep feature fusion, thereby enhancing security. To identify a face, we first locally enhance the face using a Gabor filter to capture local face information and then extract deep features using ResNet50-based deep learning architecture. The extracted features are pruned using Principal Component Analysis (PCA). On the other hand, Gait Energy Images (GEI) are used to present a gait sequence, and a convolutional neural network with a bidirectional long short-term memory (CNN-BiLSTM) is employed to capture significant gait features. PCA is applied to remove feature redundancies. After that, both face and gait features are combined using an adaptive attention model that weights the relative importance of each modality, and final classification is performed with a Softmax classifier. It is demonstrated that our final model converges well within a limited number of training iterations, achieving very high training accuracy and good validation accuracy, with clear class separability as observed in the confusion matrix. We also evaluate the performance of our final system using biometric performance curves (ROC), FAR, FRR, EER, and Grad-CAM visualization to assess whether the model is capturing the important biometric regions. Our proposed model outperforms both traditional unimodal and multimodal biometric identification methods. The proposed system can be used to securely identify humans for surveillance and authentication within and across organizations. The proposed model provides a balanced trade-off between accuracy, robustness, and interpretability in multimodal biometric recognition.

Matching contactless fingerprint images with traditional contact-based impressions has become increasingly important, especially due to the COVID-19 pandemic. Contactless methods provide better hygiene and benefit from the availability of affordable mobile phones capable of capturing high-resolution fingerprints. Traditional minutiae-based matching techniques are susceptible to errors caused by false or missing minutiae points in low-quality images, emphasizing the need for alternative features. This study explores the magnitude spectrum, a feature derived from the Discrete Fourier Transform (DFT) for matching contactless and contact-based fingerprints. A 256-bin histogram of the magnitude spectrum is generated to estimate the correlation distance to distinguish genuine from imposter attempts. Using this dataset, a Support Vector Machine (SVM) model is carefully trained and tested through 10-fold cross-validation. The results demonstrate a high matching accuracy of 97.97%, with an Equal Error Rate (EER) of 2.02% and a Rank 1 accuracy of 93.38%. The SVM classifier also achieves 96.96% accuracy in differentiating between 'genuine' and 'imposter' classes.