Energy consumption in data centers remains a critical concern for sustainable IT operations. This study analyzes energy usage patterns in the Data Center at Montclair State University, a Higher Education Institution in United States. It uses real-time environmental and power consumption data collected by a sensor network created as part of the Green IT Project, a previously funded campus project. The dataset includes temperature, humidity, energy supply and demand, and system performance metrics. Data were ingested, stored, and processed using a combination of database technologies and data visualization tools. Graphs and queries were generated to identify trends, correlations, and potential inefficiencies in the system’s energy behavior. The approach enables continuous updates and dynamic interaction with the underlying datasets. Preliminary results indicate strong correlations between environmental conditions and energy demand, with visualizations revealing specific patterns of inefficiency. These findings demonstrate the potential of data-driven monitoring to inform energy optimization strategies in institutional data centers.
Recent advances in natural language processing (NLP), particularly large language models (LLMs), have motivated the automatic translation of natural language statements into formal logic without human intervention. This enables automated reasoning and facilitates debugging, finding loop invariants, and adhering to specifications in software systems. However, hallucinations-incorrect outputs generated by LLMs are challenging, particularly for logical translation tasks requiring precision. This work introduces a novel framework that inputs English sentences, converts them into logical expressions, and then translates them into Conjunctive Normal Form (CNF) for satisfiability solving. It employs classical NLP techniques with self-defined grammar, symbolic computation libraries, and a fine-tuned language model to reduce hallucinations. In the early experiments, we observed that the fine-tuned model, trained on different grammar settings, could intentionally correct the same types of hallucinations made by the original model. Thus, it provides reliable CNF generation.
Preserving privacy is an undeniable benefit to users online. However, this benefit (unfortunately) also extends to those who conduct cyber attacks and other types of malfeasance. In this work, we consider the scenario in which Privacy Preserving Technologies (PPTs) have been used to obfuscate users who are communicating online with ill intentions. We present a novel methodology that is effective at deobfuscating such sources by synthesizing measurements from key locations along protocol transaction paths. Our approach links online personas with their origin IP addresses based on a Pattern of Life (PoL) analysis, and is successful even when different PPTs are used. We show that, when monitoring in the correct places on the Internet, DNS over HTTPS (DoH) and DNS over TLS (DoT) can be deobfuscated with up to 100% accuracy, when they are the only privacy-preserving technologies used. Our evaluation used multiple simulated monitoring points and communications are sampled from an actual multiyear-long social network message board to replay actual user behavior. Our evaluation compared plain old DNS, DoH, DoT, and VPN in order to quantify their relative privacy-preserving abilities and provide recommendations for where ideal monitoring vantage points would be in the Internet to achieve the best performance. To illustrate the utility of our methodology, we created a proof-of-concept cybersecurity analyst dashboard (with backend processing infrastructure) that uses a search engine interface to allow analysts to deobfuscate sources based on observed screen names and by providing packet captures from subsets of vantage points.
In medical chatbots and other healthcare domains, government agencies and other essential industries use sentiment analysis for public safety, improving the service quality of products, and other necessary purposes. Sometimes, the data is noisy, ambiguous, and contains mixed sentiments, where different parts of the text may convey different emotions, creating a problem in adequately detecting the sentiment. We proposed a novel recurrent neural network called the Bayesian-Lipschitz RNN with an attention mechanism to address these challenges. The Bayesian approach manages noisy and ambiguous data by representing uncertainty and avoiding overfitting. At the same time, the attention mechanism focuses on the most relevant parts of the text, effectively capturing context and mixed sentiments. Then, the Lipschitz Recurrent Neural Networks (LRNN), based on hidden-to-hidden matrices, are employed and constructed using a symmetric-skew decomposition scheme. The hidden-to-hidden matrices are used to update the hidden state and this hidden state is used to predict the sentiment, mitigate the vanishing gradient and stability issues, and got accuracy with 92.45 %.
Sentiment analysis has emerged as an indispensable resource for gleaning invaluable insights from the copious amounts of textual data produced by customer reviews, social media, and feedback forms. This research proposes a novel methodology for sentiment analysis that integrates sophisticated natural language processing (NLP) strategies with cutting-edge machine learning models to enhance the precision of sentiment categorization. The proposed methodology addresses the limitations of traditional approaches by leveraging logistic regression and Term Frequency-Inverse Document Frequency (TF -IDF) feature extraction to categorize textual sentiment into positive, negative, or neutral categories. The model is trained and evaluated using the 2023 Amazon Product Dataset, specifically focusing on software reviews. The study compares the performance of the proposed model with established sentiment analysis tools like VADER and TextBlob. Additionally, the research highlights the ethical considerations and potential biases associated with sentiment analysis in product reviews and customer feedback. The findings emphasize the importance of developing context-aware and domain-specific sentiment analysis models while addressing privacy concerns and establishing ethical guidelines for responsible implementation. This study contributes to the growing field of sentiment analysis by presenting a robust and accurate methodology that can be applied across various industries to gain a deeper understanding of customer opinions and inform decision-making processes.
With growing public health concerns related to conditions such as anxiety and depression, the need for modern technology to support affected individuals has become increasingly important. This paper details the development of a companion robot that leverages edge computing for real-time emotion recognition. Our approach utilizes a Convolutional Neural Network (CNN), fine-tuned from the ResNet-50 base model, to identify five distinct emotions: Happy, Neutral, Sad, Angry, and Fear. Video frames are captured using a Raspberry Pi camera and sent to the edge device, where face detection and emotion inference are performed. Once the inference and analysis are complete, the results are sent back to the Raspberry Pi. Device communication is facilitated through the HTTP protocol. Our model achieved a validation accuracy of 80% while maintaining an average communication time of 76.1 ms, including the duration for model inference.
Satellite Internet technology is currently booming. Due to the high cost of satellites and the limited number of orbits, a satellite must cover a large area and serve more users than before. Given that the random access process is the initial step in establishing communication between satellites and terrestrial users, congestion control has emerged as a focal point of research. In this paper, we address the random access congestion control problem in satellite Internet scenarios and introduce the soft actor-critic (SAC) method to dynamically set the access class barring (ACB) factors. The throughput of random access is optimized to allow a single satellite to support a greater number of users without experiencing congestion. The SAC-ACB method is verified by simulation to converge faster and perform better than the proximal policy optimization (PPO) method.
Lately, Deep learning and machine learning models have been vividly used in image classification to measure the accuracy of the proposed model. Deep learning networks always require high computing networks to increase the model's performance. Computer vision plays a vital role in polishing the dataset by removing unwanted information, without the loss of important features. As per certain applications, two machine learning and deep learning algorithms with different system requirements are compared and summarized. In this paper, instead of the RGB image dataset, the thermogram image dataset has been considered for the process. The main concern in considering thermal images is due to the particular application. Here the thermal images have been collected during the nocturnal hours, especially on the dawn and dusk of the day. Convolutional neural networks are the machine learning models used in CPU and GPU-cluster configurations. AlexNet will be the deep learning model used here to compare with the CNN supervised learning algorithm. Computer vision technique, HOG is used to reduce the complexity and intricacy of the dataset before getting into the network models. Because of the usage of deep network models' performance computing is expended to increase the speed of the whole process. Deer will be the spotted animal for most road accidents during the nocturnal hours in the United States. The results gaudily conclude the classification of machine learning and deep learning model detection for thermal and HOG images with the highest accuracy of 92% for the AlexNet.
In the dynamic field of 5G network monitoring, the ability to generalize monitors' placement across a network is crucial for comprehensive coverage. This study introduces the use of the Relational Graph Convolutional Network (RGCN) model to meet this challenge. We conducted a comparative analysis between the RGCN and a traditional Neural Network (NN) across two network topologies, considering every possible node configuration within these topologies. Our findings indicate that the RGCN model, once trained on a specific node pair, exhibits superior generalization ability and accuracy. It consistently transfers its learning to changes in monitors' placement and accurately estimates links delays beyond its initial training monitors. Unlike the NN, which showed significant limitations in generalizing monitors' placement and high error rates. This paper not only demonstrates the effectiveness of the RGCN model in generalizing the monitors' placement problem but also paves the way for its broader application in dynamic network monitorinf contexts.
CircularX is a compelling blockchain solution, leveraging a decentralized physical infrastructure network (DePIN) to empower businesses to turn waste into regenerative resources. Building on the foundation concept laid by CircularX 1.0 - a solution characterized by its permissioned blockchain with five key components including asset tokenization, trusted transportation, reliable valuation of regenerative waste, smart waste trading marketplace, and remanufacturing of new products, CircularX 2.0 is redesigned with a permissionless public blockchain [1]. This evolution does not only simplify the circularity model but also introduces a CircularX token to incentivize the ecosystem's participants towards more sustainable use of regenerative materials. It increases the circularity of the economy, reduces carbon footprint, extends lifespan, and produces revenue from regenerative raw materials that are otherwise treated as waste such as used coffee grounds. The architecture underpinning CircularX 2.0 integrates: 1) waste asset tokenization with a CircularX tokens reward for participants of selling, processing, and buying the raw materials and turning them into regenerative products; 2) a smart contracts-driven trading marketplace; and 3) aggregating supply and demand in localized communities to ultimately build the world's largest decentralized physical infrastructure network of regenerative resources.
In this paper, we introduce NeuroDual, a hybrid Boolean satisfiability (SAT) solver architecture that integrates Graph Attention Networks (GATs) into the Conflict-Driven Clause Learning (CDCL) process. Unlike traditional SAT solvers with fixed decision heuristics, NeuroDual leverages GATs to dynamically learn a decision heuristic specific to each SAT instance by computing an assignment score for each variable in the problem. Also, GATs in the context of SAT-solving allow NeuroDual to better understand the spatial relations of the dynamic features within a CNF clause and make informed predictions for variable assignments in the SAT instance that align more closely with the current state of the problem. For our CDCL component of NeuroDual, we implement MINISAT as the baseline solver. Our results show that incorporating machine learning techniques, specifically GATs, into SAT-solving algorithms as a decision heuristic has the potential to increase solver efficiency. These findings indicate that integrating machine learning techniques with traditional SAT-solving algorithms like CDCL to enhance the decision heuristic has the potential to drive efficiency improvements, paving the way for smarter decision-making, reduced conflict occurrences, and expedited problem resolution.
Cybersecurity has become a significant concern for automotive manufacturers as modern cars increasingly incorporate electronic components. Electronic Control Units (ECUs) have evolved to become the central control units for critical car functions such as engines and brakes, experiencing rapid technological advancements. However, this swift progression in ECU technology has also made them prime targets for cyber attacks. This vulnerability has spurred researchers to focus on securing ECUs. Numerous studies have proposed intrusion detection systems (IDS) to protect against attacks on ECUs in vehicles. Yet, these IDSs are not impenetrable; attackers can exploit them by launching evasion attacks, which can trigger numerous false positive alarms. Such false alarms can be disruptive and potentially hazardous for drivers. Additionally, attackers can evade IDSs from detecting malicious data that can cause harm to the vehicle. Accordingly, in this paper, we propose a novel training framework to train a robust in-vehicle IDS that can encounter evasion attacks. Our methodology is based on implementing two rounds of mimic learning technique for training Random Forest (RF) based IDS. RF has been chosen to incorporate the randomness of the RF architecture to enhance the robustness of the model. Additionally, in each round of the two rounds of the mimic learning technique, an RF model with a different architecture is chosen to improve the resilience of the model against evasion attacks without affecting its accuracy. Our Experimental results have shown the effectiveness of our framework against evasion attacks.
As Internet of Things (IoT) technology rapidly evolves, the widespread use and diversity of IoT devices present new challenges for device identification (often called finger-printing). Nevertheless, traditional methods for identifying IoT devices face several problems. This paper presents an identification solution capable of detecting the IoT device identities by analyzing the network traffic they generate and machine learning approaches. The solution we propose is tested on three well-known IoT traffic datasets, and showed higher prediction performance in identifying IoT device types. It is also compared against existing solutions and showed far better results, in terms of both prediction accuracy and temporal complexity.
The latest advancements in blockchain technology have significantly influenced several sectors, such as banking, healthcare, and supply chain networks. Because of its distinct attributes, like decentralization, trustworthiness, and security, this technology has recently been employed in the field of education. While the blockchain is bright, several problems, such as legal, immutability, and scalability questions, continue to hinder its adoption in education. In this paper, we look at educational application system in terms of blockchain-based solutions that have been created for educational purposes, as well as the potential challenges of blockchain technology in the field of education.
Machine learning has been successfully used to predict sentiment for general-purpose short-text sentences with relative accuracy. However, challenges around predicting legislation acceptance call for deeper analysis of longer, more interconnected text paragraphs. The E-waste management domain has a pressing need for insights into public discourse on legislation. This work presents a comprehensive approach to analyzing public discourse surrounding e-waste legislation by integrating social media data from Twitter and legislative data from LegiScan. Graph Machine Learning (GraphML) is utilized to build the prediction engine. A heterogeneous graph is constructed to capture the relationships between tweets and legislative bills using API-based data collection, preprocessing techniques, and pre-trained sentiment models. The ML pipeline includes robust preprocessing of text data, feature extraction using RoBERTa embeddings, dimensionality reduction via PCA, and the use of various Graph Neural Network (GNN) architectures (Graph Convolutional Network (GCN), Graph Attention Network (GAT), and GraphSAGE) for model training and evaluation. The final model achieved a test accuracy of 0.777, a test F1-score of 0.745, and a test ROC AUC of 0.819, demonstrating the efficacy of graph-based sentiment analysis in this context. This work's findings have significant implications for policymakers, stakeholders, and researchers in the e-waste management domain. Policymakers can use the model's sentiment analysis insights to gauge public support for specific legislative actions, providing informed and responsive decision-making. E-waste management Stakeholders can understand public opinion around e-waste legislation and develop effective communication strategies. Additionally, the proposed methodology provides an end-to-end framework for researchers to explore further applications of graph-based sentiment analysis.
Distributed Sagnac loop (SI) is a simple vibration-based sensing system, which is cost-effective and simple to implement. Because of its simplicity, its performance is low compared to other complex structure sensors. Recently, machine learning (ML) has gained significant popularity and found applications in almost each science discipline. ML is an artificial intelligence sub-field, which intimates human intelligence. Sensing is one area that has exploited ML techniques to improve and enhance the reliability of sensing systems. Using ML to enhance the reliability of location identification in SI sensing systems was proposed using frequency null as a feature for training the ML model. Extracting frequency null features for ML training degrades the system performance because of their lower power, especially under harsh environments that are very noisy. In this work, we propose using an ML model for improving event localization in SI sensing systems by exploiting time-based features. A training set containing 190 events is generated over 9.5 km of effective sensing fiber. Then, the training features are extracted using an auto-encoder neural network, which reduces dimensionality and performs denoising. After that, an ML model is developed using a random forest (RF) algorithm for predicting the event location. The obtained results indicate that the mean absolute error (MAE) of 13.4 m over 9.5 km effective sensing fiber. The percentage of samples with an MAE > 50 m is 0.52%, which is very low.
In the context of disaster management, Software-Defined Internet of Things (SD-IoT) networks playa crucial role by utilizing critical real-time data essential for addressing disasters. However, managing these networks becomes more complicated due to the heterogeneity of traffic and the load generated by each switch when transmitting information to the controller. This heterogeneity makes the controller placement issue critical. To address this challenge, we propose an opti-mization method called HTF -CP (Heterogeneous Traffic Flow-Based Controller Placement), based on the well-known Lidy+ method, which optimizes the placement of controllers while managing heterogeneous traffic in the network. We demonstrate the approach's efficiency through simulations on datasets from the Internet Topology Zoo and OS3E. The proposed method is evaluated in terms of maximum latency, average latency, and load rate. The experimental results show that our HTF -CP method provides better performance compared to Lidy+ and the random method, where the controller is selected randomly.
A digital operations management platform consolidates many data sources, enhances organisational processes, and employs analytics to foster creativity and flexibility within the organisation. To effectively handle the large number, speed, and diversity of system events that happen every day, IT requires a digital operations management platform that can minimise repetitive notifications, determine the immediate cause(s) of an IT outage, and guarantee satisfying user interactions through the application of AI/ML. AI has been highly influential in various sectors, such as manufacturing and logistics, in recent years. Manufacturing and logistics organisations are utilising AIOps (Artificial Intelligence for IT Operations) to enhance their productivity. Due to the increasing complexity of user requirements, enterprises are utilising data-driven platforms to meet client expectations. Autonomous supply chains are necessary as a result of heightened competitiveness and the growing expectations of customers. AIOps analytics platforms can assist companies in managing the heightened customer demand. This article provides a comprehensive summary of the journey and the possible advantages it offers. It also explains the necessary components for AIOps and outlines a plan for initiating the process.
Water Sort Puzzle, similar to Color Sort Puzzle, offers an engaging and addictive gaming experience where players aim to match colors in each vial. While easy to grasp, mastering the game requires strategic thinking, often necessitating players to backtrack and iterate through multiple attempts, especially at higher levels. In this paper, we delve into the application of Reinforcement Learning (RL) to conquer the Water Sort Puzzle. Our approach includes investigating Deep Qlearning (DQN) and Double Deep Qlearning (DDQN) techniques. The trained models found optimal solutions for configurations with a smaller number of vials, such as those with 5 vials. However, for more complex setups, such as those with 14 vials, although our approach outperformed random policies, further training is necessary to achieve optimal performance in terms of minimizing the number of actions required to solve the game. For a number of vials of 14, it performed better than the random policy but needs more training to win the game with an optimal number of actions.
The exponential growth in the number of connected devices within the 5G network, characterized by its high- speed, low latency, and high capacity capabilities, has presented various challenges for indoor communication. As the latest generation of mobile networks, 5G offers significant advancements over its predecessors, enabling a wide range of applications and services that require seamless connectivity and efficient data transmission. However, the characteristics of 5G network, while advantageous, also introduce unique challenges in indoor environments. One such challenge is the increased density of devices that 5G network can support. With the ability to connect a massive number of devices per square kilometer in indoor spaces. This density of devices creates congestion, leading to interference, packet loss, and delays in indoor communication. To address these challenges, our proposed solution involve a Device to Device (D2D) communication and reinforcement learning, specifically, employing Q learning. Our solution aims to make effective use of direct device communication to reduce congestion and enhance network efficiency in dense indoor environments. The successful implementation and performance evaluation of our solution hold significant implications for enhancing network performance in the context of 5G networks, particularly in dense indoor environments. This work contributes to advancing the state of efficient and reliable indoor communication by effectively addressing congestion and optimizing throughput and latency.
