The rapid expansion of the Internet of Things (IoT) has transformed modern connectivity, allowing seamless communication and data exchange between devices and systems. Nevertheless, with this increased interconnectivity comes significant cybersecurity problems, subjecting IoT infrastructures to diverse and complex cyber attacks. Therefore, developing smart intrusion detection mechanisms is now essential to ensure the integrity of data, privacy, and network trustworthiness in IoT settings.The deep learning models Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) are evaluated in this study with respect to their capacity to detect IoT cyberattacks. Three new datasets—NF-ToN-IoT, UNSW_NB15, and BoT-IoT—were employed with the preprocessing of missing value treatment, encoding categorical variables, normalization of features, and a 70/30 train-test data split. The outcome reveals excellent performance in both models with 100% accuracy on the BoT-IoT dataset. For UNSW_NB15 data, the accuracy of GRU was 97% compared to LSTM's 96%, while LSTM (83%) was slightly better than GRU (81%) for NF-ToN-IoT. These outcomes signify the stronger ability of recurrent models to handle the complexity of IoT data and strengthen the argument that model selection is based on specific features of datasets. It should be explored in future research using hybrid and transformer-based architectures to enhance emerging threat detection. Further to this, this work offers immense educational importance by presenting a practical guide towards educating students on applying deep models such as LSTM and GRU to secure IoT systems using empirical evidence and experiments in the laboratory.
Probability distributions are essential statistical instruments, which make it possible to model and analyze some probability events in various spheres, including engineering, medicine, finance, and environmental science. Classical distributions, however, have been observed to have constraints in describing the complexity of real-life data. This paper will discuss these shortcomings by proposing the DUS-transformed Generalized Polynomial Quadratic Failure Rate (DUS-GPQF) distribution, which is a new extension of the generalized linear failure rate (GLF) distribution via DUS transformation method. The DUS-GPQF distribution increases the flexibility in the GLF model which provides it with greater ability to support a larger spectrum of data behaviors. The DUS-GPQF distribution has important statistical properties which include the hazard rate functional, moments, incomplete moments, entropy and the extropy. The DUS-GPQF distribution has seventeen estimation methods, which guarantee that it is practical to apply. The tests on the DUS-GPQF distribution are performed on two real-world data sets, and the results show that the model is better than other competitive models in goodness of fit and predictive accuracy. The study offers a powerful statistical model to a complex data so as to improve theoretical as well as practical statistical analysis.
The malicious domains and adversarially crafted URLs in cyber threats evolve at a very high speed. The detection frameworks need to be robust, adaptive, and scalable in such scenarios. Traditional detection mechanisms are static feature-based approaches that cannot perform well against unseen threats, adversarial manipulations, and long-term attack evolutions. Existing systems lack granular threat attribution, cross-organization intelligence sharing, and adversarial robustness, making them unsuitable for modern cyber defenses. To address these limitations, we introduce an LLM-Based Frequent Monitoring Framework that combines five advanced techniques: Meta-Learned Self-Supervised Domain Generalization (ML-SSDG), Reinforcement Learning-Augmented Adversarial Training (RL-AdvTrain), Hierarchical Multi-Task Threat Classification (HMT-TC), Temporal Memory-Augmented Transformer for Sequential Threat Detection (TMAT-STD), and Federated Privacy-Preserving Threat Intelligence Learning (FPPTIL). ML-SSDG can achieve zero-shot detection for novel attack domains with a reduction of false negatives by 20% and enhancement of zero-shot accuracy by up to 30%. RL-AdvTrain strengthens the model against masked malicious URLs, detecting 40% more adversarial threats. HMT-TC improves threat attribution and increases classification accuracy for attacks by 50%. TMAT-STD allows for identifying emerging domain threats in real-time, while such detection reduces the response time to domain-based malware campaigns by 30%. The last one is FPPTIL, which allows shared cross-organization threat intelligence without sharing the source of private data. The process of global threat detection improves by 30%. Our framework achieves a holistic, real-time, and privacy-preserving cyber defense solution that adequately outperforms traditional approaches in adversarial resilience, threat attribution, and zero-shot detections. Taken together, these improve the cybersecurity posture, reduce false positives, and support proactive mitigation of emerging cyber threats at scale in process.
In this paper, we consider the prediction problem of the future records based on observed data from two-parameter, shape and scale parameter, Kies distribution. Different point predictors including maximum likelihood, conditional median, best unbiased and Bayesian predictors of the future records are obtained. The corresponding prediction intervals using pivotal quantity, Highest Conditional Density (HCD), Shortest Length and Bayesian prediction intervals are also developed. The Monte Carlo algorithm is used to compute simulation consistent Bayesian prediction intervals for future unobserved records. The performance of the so obtained point predictors and prediction intervals are compared via experimental numerical simulation. The criteria that were considered for comparison purposes are mean square prediction error (MSPE) and prediction bias for point predictors and coverage probability (CP) and the average length (AL) for prediction intervals. A real and simulated data sets are performed for illustrative purposes.
Modern uncertainty-modeling frameworks-fuzzy sets, intuitionistic fuzzy sets, hyperfuzzy sets, neutrosophicsets, and plithogenic sets-provide powerful tools for capturing vagueness and imprecision. In particular, neutrosophic sets characterize each element by three independent degrees: truth, indeterminacy, and falsity. Classical neutrosophic sets have been refined by partitioning the membership degrees into additional components. Recently, the Hexapartitioned Neutrosophic Set, Octapartitioned Neutrosophic Set, Nonapartitioned Neutrosophic Set, and Decapartitioned Neutrosophic Set have been defined. In this paper, we introduce four novel partitioned models-hexapartitioned, octapartitioned, nonapartitioned, and decapartitioned neutrosophic offsets/oversets/undersets-and show how each can be seamlesslyembed ded into the plithogenic-set framework.
