The escalating complexity of cyber threats requires increasingly sophisticated approaches to automated detection, particularly within the domain of ransomware. Existing methods, such as signature-based and heuristic detection, often fail to recognize novel ransomware variants due to their reliance on predefined, static indicators or anomaly-based heuristics, leading to high false positive and false negative rates. DeepCodeLock addresses these limitations through a deep learning framework that leverages convolutional and recurrent neural network layers to analyze ransomware behavioral patterns comprehensively, allowing for the extraction of high-level features indicative of malicious intent. By processing system activities like file access, network connections, and process hierarchy structures, the model autonomously generates a dynamic, composite behavioral signature for each monitored activity, enhancing detection precision. Experimental results demonstrated that DeepCodeLock achieved significant improvements in detection accuracy, reduced false positive and negative rates, and maintained low latency and resource utilization across varied network sizes, establishing its practicality for real-time, large-scale cybersecurity applications. Overall, DeepCodeLock's methodological innovations emphasize its utility as an adaptive and scalable ransomware detection system, offering new possibilities in protecting critical infrastructures from emergent cyber threats.

We analytically demonstrate how the properties of local cascaded metasurface pairs can be mapped to those of single non-local metasurfaces such that their external field scattering remains identical regardless of the particular incident field. This mapping is obtained for two-dimensional problems by exploiting Love's equivalence principle and fundamental linear operations. We verify the mapping with simulated data and show that the mapping preserves the losslessness, passivity, and reciprocity characteristics of the original local cascaded metasurface pair.

This paper presents our participation in the Capsule Vision 2024 Challenge, where we leverage FasterViT, a high-speed variant of the Vision Transformer (ViT), to address the problem of efficient image classification under resource constraints. By adapting FasterViT with targeted data augmentation, fine-tuning, and optimization strategies, we achieved improved accuracy and computational efficiency over standard baselines. Our approach demonstrates the potential of efficient transformers in real-world visual recognition applications.

AI systems increasingly shape critical decisions across personal and societal domains. While empirical risk minimization (ERM) drives much of the AI's success, it typically prioritizes accuracy over trustworthiness, often resulting in biases, opacity, and other adverse effects. This paper discusses how key requirements for trustworthy AI can be translated into design choices for the components of ERM. We hope to provide actionable guidance for building AI systems that meet emerging standards for trustworthiness of AI.

Current technology utilizes a sophisticated standard room measurement system to analyze the average roughness of stainless steel bearing surface. However, this process relies on random sampling at infrequent intervals, resulting in inefficiency and inadequate analysis of a large number of samples. To address this issue, we propose a solution using computer vision via laser scanning. The scanning process produces an amplified speckle pattern of the ring surface. This pattern is then analyzed to verify the number of edges that cross a predefined threshold and exhibit symmetrical patterns, particularly on smoothed ring surfaces. Remarkably, this system operates with a turnaround time of less than a second. Moreover, unlike the standard room setup, this technology can be seamlessly integrated directly into production/manufacturing units.

Dynamic Behavioral Profiling (DBP) presents a novel approach for enhancing ransomware detection through continuous, adaptive monitoring of system behaviors, effectively addressing limitations inherent in static and heuristic-based methods. DBP operates through a multi-stage system architecture, enabling real-time analysis of deviations from baseline behaviors, which are indicative of ransomware activities across various operational phases. Employing techniques such as adaptive thresholding, entropy-based analysis, and time-series modeling, DBP identifies anomalies associated with encryption processes, high-frequency file modifications, and system anomalies linked to ransomware, providing a robust mechanism for early threat detection. Extensive evaluations of DBP reveal significant improvements in detection accuracy, particularly against polymorphic variants, achieving a high level of resilience in identifying unknown threats that evade signature-based methods. Furthermore, DBP demonstrates low false-positive rates, achieving a balance between sensitivity to ransomware indicators and specificity for benign anomalies. By managing computational overhead and optimizing resource consumption across different network conditions, DBP maintains operational efficiency suitable for largescale implementation in diverse cybersecurity environments. Through an analysis of DBP's performance metrics-detection accuracy, false positives, latency, and resource utilization-the results substantiate DBP's applicability as a scalable, effective solution in mitigating ransomware threats within complex and dynamic digital infrastructures.

Physical layer key generation from the wireless channel is an emerging area of interest to provide confidentiality and authentication. One of the main challenges in this domain is to increase the length of the secret key while maintaining its randomness and uniformity. In this work, new dimensions for wireless channel-based key generation are proposed for orthogonal frequency division multiplexing (OFDM) systems. The novel perspective of the proposed scheme called NIAKG lies in the generation of key bits not only from the magnitudes of OFDM subchannels as it has conventionally been done but also from the number and positions/indices of those subchannels, whose channel gains are above the mean of the respective subblock. The effectiveness of the proposed algorithms is evaluated in terms of key generation rate and key mismatch rate. It is shown in the simulation results that the involvement of the proposed dimensions can double the key generation rate compared to conventional algorithms.

In this paper, we introduce ZIMUX, a novel Non-Orthogonal Multiple Access (NOMA) technology designed to achieve completely interference-free, secure, and reliable transmission in future wireless networks (6G and Beyond). Unlike conventional NOMA, which struggles with increased complexity and inter-user interference, ZIMUX leverages the design of a novel type of frequency domain-based precoders to eliminate inter-user interference, while ensuring simplified receiver operations. Key benefits of ZIMUX include reduced receiver complexity, lower transmission latency, decreased signaling overhead, and enhanced security against both external and internal eavesdropping. Our design enables the seamless combination of users regardless of their distance from the base station, making it adaptable to a wide range of communication scenarios. The simplicity and efficiency of ZIMUX position it as a candidate solution for applications requiring low complexity, such as the Internet of Things (IoT), while maintaining high levels of security and performance.

Objective: Inertial sensors have the potential to be a useful clinical tool because they can facilitate human motion capture outside the research setting. A major barrier to the widespread application of inertial motion capture is the lack of accepted calibration methods for ensuring accuracy, in particular the lack of a common convention for calculating the rotational offset of the sensors, known as sensor-to-segment calibration. The purpose of this study was to develop and test a sensor-to-segment calibration method for upper limb motion capture which is practical for clinical applications. Methods: We developed a calibration method which depends mainly on the estimation of joint axes from arbitrary elbow motion, and partially on the design of custom attachment mounts to achieve physical alignment. With twenty healthy participants, we used OpenSim's inertial sensor workflow to calculate joint kinematics, and evaluated the accuracy of the method through comparison with optical motion capture. Results: We found the new calibration method resulted in upper limb kinematics with a median RMS error of 5-8°, and a median correlation coefficient of 0.977-0.987, which was significantly more accurate than a pose-based calibration (p-value < 0.001). Conclusion: This work has demonstrated a method of calibration which is practical for clinical applications because it is quick to perform and does not depend on the subject's ability to perform specific movements, or on the operator's ability to carefully place sensors. Clinical impact: The calibration method proposed in this work is a realistic option for the translation of inertial sensor technology into everyday clinical use.

Document classification is an essential task for many Cyber Threat Intelligence (CTI) applications that source opensource documents to identify cyber threat patterns such as Indicators of Compromise (IoC), TTP (tactics, techniques, and procedures), Named Entities and more. Traditional machine learning systems require a large number of labelled datasets to automate the document classification process, which is not suitable in the ever-changing CTI landscape. The recent advancements in various Pre-trained Language Models (PLMs) have motivated the popularity of zero-shot text classification approaches. However, a less-studied problem is how such zero-shot methods can be effectively applied in the cybersecurity domain and how their performance is on the document classification task. In this paper, we introduce a self-training approach named Latent Space Refinement (LSR) for zero-shot document classification for CTI applications. Experiments show that our LSR-based method outperforms the relevant state-of-the-art baselines on various scopes of document classification tasks significantly.

The instability of electrical trading prices presents considerable challenges for forecasting and allocating resources in smart grid systems. This paper proposes an electrical price forecasting approach based on hybrid stacking ensemble learning methods. Although weather conditions significantly impact the distribution of electrical loads and prices, different weatherrelated historical data, including temperature and humidity, would be utilized as input features for the proposed hybrid stacking model based on different Artificial Intelligence (AI) methods. Different from the conventional load forecasting method, the proposed hybrid stacking method includes multiple AI models including XGBoost, CatBoost, Neural Oblivious Decision Ensemble (NODE), Light Gradient-Boosting Machine (LightGBM), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM), with Lasso regression acting as the meta-learner. Considering distributed weather data inputs and the stacking nature of the proposed method, the model seeks to improve forecasting accuracy and more closely represent actual electrical load and price situations. The performance of the hybrid stacking model was experimented with realistic datasets and evaluated by different time series forecasting metrics, including Mean Absolute Percentage Error (MAPE). Comparisons with single AI models show that the proposed hybrid stacking method improves the performance of electrical load and price forecasting and could be beneficial to the operation and allocation of smart grid systems.

Cyberattacks leveraging encryption-based extortion techniques have grown both in frequency and complexity, creating an urgent need for more advanced detection mechanisms capable of identifying malicious activity before significant harm occurs. The Dynamic Encryption Pattern Analysis (DEPA) system is introduced as a novel solution to address this challenge, focusing on the detection of ransomware through real-time monitoring of encryption behaviors. Unlike traditional methods that rely on static signatures or general anomaly detection, DEPA analyzes file modification patterns and entropy changes, allowing it to identify ransomware activity at an early stage. The system's adaptive detection algorithm further enhances its effectiveness, adjusting to evolving ransomware techniques through continuous feedback loops. Extensive experiments were conducted using a diverse dataset of ransomware families, demonstrating DEPA's high detection accuracy, low false positive rates, and efficient resource management across multiple operating systems and environments. In addition to its technical advantages, DEPA's modular architecture supports seamless integration into various cybersecurity infrastructures, making it suitable for large-scale deployments in both enterprise and cloud-based systems. This work highlights the potential of DEPA as a robust, scalable, and efficient solution for mitigating the rising threat of encryptionbased cyberattacks, offering a promising approach to enhancing automated ransomware detection capabilities.

This study presents wideband propagation measurements of 105 GHz multipath characteristics, encompassing a full 360°in a real office desktop environment. High-speed wireless personal area network (WPAN) systems operating in such environments represent a promising use case for sub-terahertz (THz) communication systems owing to the short-range nature of such networks. Additionally, selecting a frequency band close to the millimeter-wave spectrum increases the feasibility of sub-THz WPAN systems compared to the widely recognized 300 GHz band, mainly because of the availability of low-cost hardware. However, the multipath propagation characteristics at the 105 GHz band, specifically within a 360° range in a real office desktop environment, have not been thoroughly investigated. To address this gap, we evaluate the 105 GHz multipath propagation characteristics, considering both delay and angular profiles, and compare them with our concurrent 60 GHz measurements in the same environment. The results indicate a notable distinction between the two bands: a physical partition maintaining personal space causes the multipath power at 105 GHz to deviate by 10 dB relative to the 60 GHz band. Furthermore, our system-oriented analysis highlights the similarity of propagation characteristics in both bands, as nearly all multipath waves at 105 GHz exhibit power levels comparable to those observed at 60 GHz. In both frequency bands, the delay spread extends up to 5 ns, while the angular spread reaches up to 40°. These findings suggest that the current 60 GHz WPAN system standards could be effectively extended to the 105 GHz band for sub-THz WPAN applications.

Point cloud video streaming over networks is challenging because of the extremely high data rate of uncompressed point cloud data. Adaptive point cloud video streaming has been proposed to deal with this challenge. However, temporal quality variation and stalling might occur under unstable network conditions, potentially degrading users' Quality of Experience (QoE). This paper aims to evaluate and model the impacts of temporal quality variations and stalling on users' QoE in adaptive point cloud video streaming. We first conduct a largescale subjective study to construct a QoE database. Then, based on the constructed database, the effects of individual factors are analyzed, and two novel QoE prediction models are presented. Experiment results show that the proposed QoE models achieve high prediction performance in terms of PLCC, SROCC, and RMSE across various point cloud videos.

In this paper, we develop a Virtual Reality-based immersive learning environment that allows teachers to conduct a lesson in a virtual space. The proposed system allows teachers and students to be present in the same virtual space regardless of their actual physical locations. The teachers can verbally communicate with students in real-time, sharing/collecting learning materials, and conducting in-class assignments. The students can interactively learn about real-world objects by touching and manipulating 3D models with hand tracking. Evaluation results demonstrate the effectiveness of the proposed system.

Point cloud video is an effective method to represent moving objects for metaverse applications. Real-time streaming of point cloud video can offer truly immersive experiments for metaverse users. However, point cloud videos have an extremely high data rate and demand significant processing resources for compression and rendering at the user device. To address the above challenges, this paper presents a novel server-driven viewaware point cloud video streaming system that can effectively reduce bandwidth requirements by combining hidden point removal and video-based point cloud encoding. Experiment results show that the proposed method can reduce the bandwidth requirement by up to 24% compared to the baseline method.

The increasing computational demands and inefficiencies associated with the scaling of advanced language models present significant challenges, particularly in terms of maintaining accuracy, inference speed, and resource utilization. Addressing these issues, the introduction of Dynamic Token Compression (DTC) and Adaptive Layer Pruning (ALP) offers a novel and transformative approach to optimizing model performance, significantly reducing token redundancy and selectively pruning model layers based on the complexity of input queries. These methods enable enhanced language models to generate more coherent, contextually accurate outputs while minimizing computational load, making them more suitable for real-time applications. Through comprehensive experimentation, DTC and ALP were shown to improve model efficiency across multiple key metrics, including perplexity, BLEU score, and hallucination rate, while also reducing memory usage and energy consumption. The ability to dynamically adjust token processing and layer utilization without compromising on performance demonstrates the practical value of these techniques for large-scale model deployments in various domains. By optimizing both computational and linguistic aspects of the models, this research provides a robust framework for future advancements in language model architecture.

Effective resource orchestration for network slicing is critical for optimizing the performance of diverse applications running on next generation communication networks. This paper presents a novel approach that leverages advancements in multiagent reinforcement learning (MARL) to adaptively learn the resource requirements of various applications in network slices and orchestrate resources in real-time. Our proposed MARL-based orchestration scheme aims to balance the varying requirements of individual network slices, ensuring optimal performance amid dynamic application deployments with limited network information. Simulation results and comparative analyses validate the efficiency and efficacy of our methodology, demonstrating its superiority over traditional methods in terms of system performance and resource utilization. Simulation results indicate that our strategy significantly enhances system utility and efficiency, particularly with limited resources.