Park College of Engineering and Technology

The growing complexity of cyber-attacks and distributed computing environments has exposed limitations in traditional intrusion detection systems, including privacy concerns, scalability issues, and limited explainability. This paper proposes the Federated Hierarchical Attention Transformer with Adaptive SHAP Explainability (FHAT-ASHAP), a privacy-preserving framework for cybersecurity threat detection. The model combines hierarchical attention mechanisms and transformer networks to learn packet-, flow-, and session-level traffic patterns within a federated learning environment. A trustaware aggregation strategy enhances robustness and communication efficiency across distributed clients. To improve transparency, Adaptive SHAP Explainability integrates attention-based learning with feature attribution analysis, providing interpretable threat predictions. Additionally, an autoencoder-based anomaly detection module identifies previously unseen and zero-day attacks. Experimental evaluation on CICIDS2017, CSE-CIC-IDS2018, and UNSW-NB15 datasets demonstrates high detection accuracy, reduced false-positive rates, improved privacy preservation, and enhanced explainability. The results indicate that FHAT-ASHAP is an effective solution for next-generation intelligent cybersecurity systems.

Object detection using monocular 3D in video sequences is still a challenging task because of the depth ambiguity and temporal variations of objects due to their motion, occlusions, and a changing viewpoint. The current work suggests a Hierarchical Temporal Memory Recurrent Transformer (HTM-RTNet), a framework to jointly predict the dynamics of short-term motions and the long-term time-dependencies between a pair of consecutive frames. This framework combines most of the multi-scale spatio-temporal features extraction with a hierarchical temporal memory module that allows to track objects, depth continuity, and temporal consistency across occlusions and viewpoint changes. An anchor-free 3D detector head is a head that directly predicts the location, orientation and dimensions of the object using monocular video inputs. HTM-RTNet is coded in PyTorch and tested on general autonomous driving datasets, such as KITTI and nuScenes. Experiment findings indicate a 3784% relative to the current monocular technologies: 3784% increase in 3D AP and BEV AP, 1520 decrease in mean depth error, and more than 50 enhancements in inter-frame variance and fragmentation of the trajectory. The suggested method is effective to stabilize object trajectories in 3D and minimize depth variations between frames. The significance of hierarchical memory and recurrent transformer modules in preserving the long-term temporal coherence is supported by ablation studies. In demanding conditions of occlusions, high-speed motion as well as dynamic viewpoints, qualitative analysis shows uniform performance. HTM-RTNetackieves 3D perception which is temporally stable, accurate, and computationally efficient when monocular video is used. The findings illustrate the advantages of the memory-based temporal learning to autonomous driving in the real world. This architecture provides better reliability in dynamic scenes, by closing the range between the single frame detection and video-based 3D perception. The approach lays the groundwork to subsequent extensions of the method as using multi-camera systems or self-monitored depth estimation. In general, HTM-RTNet provides a solid, statistically confirmed method of temporally coherent, monocular object detection of 3D objects in video sequences.

The growing adoption of electric vehicles (EVs) highlights the need for efficient energy management strategies (EMS) to optimize hybrid energy storage systems (ESSs) and reduce operating costs; yet, traditional controllers struggle with nonlinear powertrain dynamics and fluctuating demand and supply conditions. To address this challenge, we propose a hybrid energy management framework that integrates a Takagi–Sugeno fuzzy controller with Portia spider optimization (PSO) and a mix style neural network (MSNN), termed PSO-MSNN. In this approach, PSO optimizes fuzzy controller parameters, while MSNN predicts optimal vehicle and system states for improved decision making. The main goal of the proposed PSO-MSNN technique is to minimize the overall operational cost of EV energy management systems, reduce energy consumption, and improve system efficiency. In this approach, PSO optimizes fuzzy controller parameters, while MSNN predicts optimal vehicle and system states for improved decision making. The method is validated in MATLAB and compared with the deep deterministic policy gradient algorithm (DDPGA), gray wolf optimization (GWO), and adaptive salp swarm algorithm (ASSA). The results show that PSO-MSNN achieves the lowest cost of 1and energy consumption of 710.20 kJ. These outcomes demonstrate that PSO-MSNN significantly enhances cost efficiency, energy utilization, and adaptability, underscoring its potential for advancing robust and scalable EV energy management systems.

The digital payment systems allow customers to make payments quickly and conveniently, while also improving access to various financial services. For financial institutions, digital payments allow them to process and offer payment services in a faster and more organized manner. However, with the increased use of digital payments, there has been an increase in fraudulent activities. This has become a major concern for customers as well financial institutions. Traditional fraud detection methods use fixed thresholds and static rules to spot suspicious behavior. Because of this, detection methods created before digital payment technologies became common generate many false positives and struggle to identify new and evolving fraud patterns. This paper introduces a real-time fraud detection framework aimed at spotting suspicious UPI transactions. The framework seeks to accurately classify both known fraud patterns and new types of fraud while also identifying unusual transaction activities that have not been labeled as fraudulent before. A hybrid machine learning approach is used, combining a supervised Random Forest (RF) algorithm with an unsupervised Isolation Forest (IF) algorithm. Both models examine each transaction on its own, producing separate risk assessments that are then combined using a weighted risk score fusion method to evaluate the chance of fraud based on transaction features and types of fraud. Experimental results show that the proposed hybrid machine learning framework is better than individual supervised and unsupervised models. It achieves higher detection accuracy and reduces false positives in real-time UPI transaction environments.

The reliability of power transformers is significantly affected by the condition of their insulating oil, which deteriorates over time due to thermal, electrical, and aging characteristics. To address this challenge, the present study examines the effectiveness of zinc oxide (ZnO) nanoparticles and hybrid ZnO-Nickel(II) oxide (NiO) nanocomposites in enhancing the performance of aged mineral oil. The ZnO materials synthesized by coprecipitation exhibit a nearly spherical morphology with an average particle size of approximately 20 nm, as confirmed by structural analysis. ZnO-NiO nanocomposites were prepared by dispersing ZnO nanoparticles in a CTABstabilized suspension, followed by the controlled reduction of nickel ions. After that, dispersion of ZnO-NiO nanomaterials in aged mineral oil yields stable nanofluids upon ultrasonication. To assess the insulation performance of these nanofluids, their electrical, thermal, and aging properties were systematically measured. The findings reveal an enhancement in the electrical strength and thermal stability of the aged oil after nanoparticle incorporation. This study confirms that, compared to ZnO, ZnO-NiO nanoparticles more effectively restore transformer oil functionality, thereby maximizing the durability and reliability of the transformer insulation.