Prime University

Robust and personalized anomaly detection is essential due to the rapid growth of UAV deployments in critical infrastructure, logistics, and surveillance. Distributed, non-IID, and sensitive UAV communication scenarios pose challenges for traditional centralized learning. To address these issues, this work presents BANCO-FL, a balanced and optimized federated learning framework combining a lightweight neural network with adaptive aggregation methods, FedAdam, FedMedian, and ClusterAvg. Experiments conducted on a real-world UAV dataset containing 2.35 million communication records demonstrate that BANCO-FL achieves a peak accuracy of 99.98%, 99.98% precision, 99.98% recall, and a 99.98% F1-score in 3-client and 9-client non-IID scenarios. Compared to standard baselines, BANCO-FL reduces misclassification rates by over 35%, improves training stability, and enhances fairness across clients. These findings show that BANCO-FL is a practical, scalable, and communication-efficient solution for real-world UAV anomaly detection.

Breast cancer is one of the leading causes of mortality among women in the world where late or false diagnosis compromises death to a large extent. Despite its promising diagnostic power, machine learning has not had high clinical credibility and usage because it cannot be interpreted, which reduces its practical usage. This paper introduces a comprehensible and explicable artificial intelligence-based ensemble model on Breast Cancer Wisconsin Dataset. Different machine learning models have been put to test such as Logistic Regression, Decision Tree, Support Vector Machine, K-Nearest Neighbors, Naive Bayes, Random Forest, Extra Trees, Gradient Boosting, AdaBoost, Bagging and XGBoost and best of them were combined in the form of the soft voting ensemble. The two most successful Logistic Regression and Support Vector machine models are combined with the soft-voting ensemble that is optimized with the use of the GridSearchCV with the ROC-AUC as the parameter of interest. The proposed ensemble had a 98.25 % accuracy, $\mathbf{1 0 0 \%}$ precision, $\mathbf{9 5. 3 5 \%}$ recall, $\mathbf{9 7. 6 2 \%}$ F1-score, and 99.74% ROC-AUC on the held-out test set. With no incorrectly-determined positives and only two incorrectly-determined negatives (Misclassification Rate: 1.75%). Novel LIME and SHAP methods were used in order to provide explainability and interpretability in diagnosis by global and local explanations. These findings confirm that explainable ensemble models have the potential to provide very precise, interpretable and clinically trustworthy diagnoses of breast cancer, which can ensure safer and more transparent decision-making in the field of women healthcare.

Intrusion detection maintains the security of the networks by identifying any attempts to attack or disrupt systems; however, the conventional signature-based techniques often fail to identify new and sophisticated attack patterns. Modern studies indicate that the machine learning methodologies have the potential of identifying threats that were previously unknown. In the paper, we improve an intrusion detection system based on the ensemble learning methods and test the framework with a publicly available dataset. We compare popular models of ensembles, such as Random Forest, Extra Trees, XGBoost, LightGBM, CatBoost, AdaBoost, and Gradient Boosting, in order to determine the most effective one. To advance the transparency of both professionals and amateurs, LIME is used to receive the model outputs, prove that protocol characteristics, transport-layer indicators, and HTTP/DNS header data play a major role in the classification. Across all tests, Random Forest consistently achieves accuracy rates exceeding 99%, coupled with low false positive rates and impressive outcomes in Precision, Recall, F1-score, Balanced Accuracy, and Cohen's $\kappa$. In summary, we present a clear, accurate, and explainable approach to reinforcing computer and network shield to protect against evolving malicious cyber activities.

The deployment of intelligent reinforcement learning (RL) agents on resource-constrained edge devices remains a fundamental challenge due to the substantial memory, computational, and energy requirements of modern deep learning systems. While large language models (LLMs) have emerged as powerful architectures for decision-making agents, their multi-billion parameter scale confines them to cloud-based deployment, raising concerns about latency, privacy, and connectivity dependence. We introduce BitRL, a framework for building RL agents using 1-bit quantized language models that enables practical on-device learning and inference under severe resource constraints. Leveraging the BitNet b1.58 architecture with ternary weights (-1, 0, +1) and an optimized inference stack, BitRL achieves 10-16x memory reduction and 3-5x energy efficiency improvements over full-precision baselines while maintaining 85-98 percent of task performance across benchmarks. We provide theoretical analysis of quantization as structured parameter perturbation, derive convergence bounds for quantized policy gradients under frozen-backbone architectures, and identify the exploration-stability trade-off in extreme quantization. Our framework systematically integrates 1-bit quantized language models with reinforcement learning for edge deployment and demonstrates effectiveness on commodity hardware.

Indoor Positioning Systems (IPS) based on Received Signal Strength Indicator(RSSI) signal measurements and Bluetooth Low Energy Beacon (BLE Beacons) have severe challenges to address. To name a few: signal interference, loss of signal packets, incomplete signal transmission, multipath fading, etc. BLE-Beacon transmitted RSSI data-based indoor positioning systems suffer from significant errors in accuracy due to environmental effects on the signals received. This study illustrates a novel strategy to increase system accuracy while not sacrificing system robustness. The approach consists of applying DBSCAN clustering to RSSI signal patterns to remove outliers that were noisy RSSI readings and then integrating these clusters with the Federated Kalman Filter (FKF) framework. To take care of such problems, this approach provides a novel hybrid method that integrates the Federated Kalman Filter (FKF) framework with RSSI clustering. With this, it utilizes the Position-Annotated BLE RSSI Dataset. The results showcase a significant increase in accuracy and robustness over traditional approaches. It also highlights the potential of a highly scalable and efficient indoor positioning solution. The study conducted in this paper results in a 13