Pacific Life

This paper examines the role of financial advisors in promoting annuity literacy in the U.S. and the moderating effect of financial knowledge on this relation-ship. Using a 2023 survey dataset from a large U.S. insurance company and employing propensity score matching, the findings show low overall annuity literacy but a positive association between financial advisor use and annuity literacy. The study further identifies financial knowledge as a moderating factor in this relationship, while also highlighting the potential for diminishing returns from financial knowledge when a financial advisor is involved. These findings highlight the importance of financial advice and knowledge in understanding annuity products. The paper also offers recommendations for financial advisors to work with clients of varying financial knowledge levels effectively. [Key words: annuities; annuity literacy; financial knowl-edge; financial literacy; financial advice.]

The success of Network Intrusion Detection Systems (NIDS) is more or less determined by the quality of datasets available that can reflect a variety of attack scenarios. Nonetheless, popular benchmarks like NSL-KDD and CICIDS2017 are characterized by a very high level of imbalance between the classes and insufficient representation of frequent but not very significant intrusions, like User-to-Root (U2R) and Remote-to-Local (R2L) attacks. In order to address these constraints, this paper suggests a Generative Adversarial Network (GAN)-based system to synthesize samples of attack to be used as an augmentation to existing datasets. Different settings such as DCGAN, WGAN and Conditional GAN (cGAN) were tested to produce a realistic artificial traffic, but still maintain the statistical distribution of real-life statistics. Experimental data prove that using GAN generated samples in the training stage greatly enhances the performance of the classifier, where recall of the minority attack classes have risen by over $50-60 \%$, with deep learning models (CNN, LSTM) scoring over 0.95 on the AUC. Moreover, cGAN was shown to be the most efficient one, resulting in class-conditional samples that boosted the detection in all the classifiers tested. Such evidence shows that the GAN-based augmentation is a feasible and efficient solution to the issue of the dataset lack and disproportions in the field of intrusion detection research with the door to more resilient, effective, and precise NIDS.

Radiographic imaging is a cornerstone of modern diagnostics, yet conventional automated systems often rely on single-task deep learning models, limiting their ability to capture the complexity of multi-disease presentations. To address this gap, this paper proposes an IoT-integrated multi-task deep learning (MTDL) framework for the simultaneous detection of multiple diseases in radiographic images. The framework employs a shared convolutional backbone with task-specific heads for disease classification, lesion localization, and severity assessment. IoTenabled imaging devices provide real-time data acquisition, edgelevel preprocessing, and secure transmission, while cloud-based inference ensures scalable computation. Experiments on NIH ChestX-ray14 and CheXpert datasets demonstrate that the proposed model achieves superior performance compared to baseline CNNs, with AUC > 0.92 for classification, mAP gains of $\sim 10 \%$ for lesion localization, and QWK improvements for severity prediction. Furthermore, IoT integration reduced latency and bandwidth consumption, enabling practical deployment in telemedicine and smart healthcare environments. The results confirm that combining IoT infrastructures with MT-DL architectures enhances diagnostic accuracy, reduces inference time, and provides interpretable outputs, thereby making the system highly suitable for real-time, resource-constrained healthcare ecosystems.

Hybrid infrastructures of clouds are becoming popular in balancing enterprise computing cost efficiency, flexibility and scalability. Nevertheless, in these heterogeneous environments, workload performance is a considerable challenge because configuration is different, workload is heterogeneous, and network conditions change dynamically. This article reports an AI-based predictive model using Gradient Boosting methods (XGBoost, LightGBM, and CatBoost) to predict the execution time, throughput, and resource consumption of hybrid workloads. The framework combines information gathering of containerized and VMbased tasks in both public and private clouds, processes features systematically previously to obtaining resourceperformance interactions, and utilises boosting models to carry out accurate forecasting. The results of experiments prove that Gradient Boosting outperforms classic regression, SVM, neural networks significantly ($\mathrm{R}^{2}>0.92$, prediction errors (MAE and RMSE) are reduced by $30-40 \%$. Moreover, the analysis of feature importance showed CPU allocation and request rate to be prevailing predictors, which is useful in scheduling a workload and meeting the SLA. These findings identify Gradient Boosting as an effective and interpretable tool to support intelligent autoscaling and resource management in hybrid cloud settings.