Industrial Internet of Things (IIoT), or Industry 4.0, is an application of IoT in the industrial sector. Its main objective is to enhance product quality and optimize production costs by leveraging advanced technologies such as edge/fog/cloud computing, 5G/6G, and artificial intelligence. In the context of Industry 4.0, numerous devices and systems are interconnected to provide seamless services to users. However, with this interconnection comes the need to protect these devices and the information they transmit from cyberthreats and intrusions. In order to tackle this challenge, our proposed solution involves the utilization of deep learning (DL) models to develop an anomaly-based detection system. Our approach involves two powerful DL models, namely Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU). The proposed model’s performance is studied within binary and multiclass classification using a new real-world industrial traffic dataset called Edge-IIoTset. The outcomes of our experiments showcased the efficacy of the CNN-GRU model that we proposed, surpassing the performance of recent related works in terms of performance metrics, including accuracy, precision, false positive rate, and detection cost. The combination of the two models CNN and GRU outperforms the GRU model with 88% of detection cost in multiclass classification for one traffic flow.
This paper presents an intelligent cyber-physical system framework for detecting and managing type 2 diabetes. The framework leverages the benefits of cloud and federated-edge computing while addressing concerns of security, delay, and communication costs. It enables the seamless integration of knowledge extracted from machine learning (ML) models trained at the edges into a CNN model using distillation and ensemble techniques. To identify the effective model for detecting diabetes, we performed a comparative study using four different ML classifiers: Support Vector Machine (SVM), Artificial Neural Network (ANN), Case-Based Reasoning with Fuzzy K-Nearest Neighbor (CBR-FKNN), and K-means with Logistic Regression (K-Means-LR) at a single edge. We employed various preprocessing techniques and feature selection algorithms to identify the optimal features for the four models. An experimental evaluation was conducted using two diabetes datasets, the PIMA dataset and the Diabetes dataset 2019. The experiments were mainly conducted to assess our proposed framework but not the datasets. The results demonstrated that the SVM model outperforms other models in diabetes detection. The results also showed that the Diabetes dataset 2019 provided better accuracy and F1 scores than the PIMA dataset.
Wormhole attack has exposed-mode (internal attack with exposed attacker nodes identities) and hidden-mode (external attack with hidden attacker nodes identities). In exposed-mode, the pair-wise connected attacker nodes fool the legitimate nodes by using a hidden link to route packets and yield high packet delivery ratio. As the packets reach the wormhole nodes, attackers can initiate traffic analysis, packet dropping, and/or packet modification attacks. This paper studies and analyzes the impact of the exposed-mode of wormhole attack in Opportunistic Mobile Networks (OMNs) and the parameters affecting it. The impacts of the exposed-mode of wormhole attack are analyzed using the amount of extra routed packets the attacker nodes will obtain. The attack was launched by varying different parameters (i.e., number of wormhole nodes, attack frequency, and attack duration) that influence its intensity against four main routing protocols in OMNs (i.e., Prophet, Spray and Wait, Epidemic, and First Contact). The simulation experiments employed two widely-used mobility traces (real-world and synthetic) in OMNs and were analyzed in terms of the most vulnerable routing protocol and the most influential attack parameter. We concluded that attackers can smartly overthrow safe communications in OMNs using deep analysis of routing mechanisms and nodes density.
Traditional methods for video recognition require hand-crafted features, which often involves offline pre-processing for real-world videos. In this study, we propose a conceptually simple framework that directly takes raw videos as an input source for activity recognition. Our framework consists of two streams, namely a spatial stream and a temporal stream. The spatial stream is trained on RepVGG-B0 ConvNet using cropped RGB features, while the temporal stream uses an attention-based Bi-directional Long Short-Term Memory (Bi-LSTM) network to learn posture vectors from human pose data obtained through Faster R-CNN pre-trained model. Our proposed method is evaluated on a standard video action recognition benchmark, MSR Daily Activity3D, and proves to be competitive with state-of-the-art action recognition methods. We achieve state-of-the-art performance on MSR Daily Activity3D with a precision and recall rate of 99.01% and 98.91%, respectively. Our results demonstrate the effectiveness of our approach in recognizing video actions.
As stated by the United Arab Emirates’s (UAE) Community Development Authority (CDA), there are around 3,065 individuals with hearing disabilities in the country. These individuals often struggle to communicate with broader society and rely on scarce sign language (SL) interpreters. Moreover, Arabic’s dialects diversity compounds the issue by causing dialects in the Arabic Sign Language (ArSL). Hence, the call for a standardized reference for ArSL in the region is a priority. To address these challenges, we’ve developed an Emirate Sign Language (ESL) electronic dictionary (e-dictionary) with a dataset of 127 signs and 50 sentences, recorded by hearing-impaired individuals in the UAE with various degrees of deafness. Supervised by certified interpreters and validated by ESL’s department head at CDA in Dubai, the recordings were made using Azure Kinect DK, resulting in 708 recordings. The dataset is then processed to 10fps. The e-dictionary offers features such as webcam-based sign recognition using YOLOv8 technology, voice-based signing via Arabic Automatic Speech Recognition, text-based signing, and words spelling in ArSL.
In order to operate effectively, the orbital parameters of the satellite are essential, to estimate satellite position over the previous and in the near future. For the planning of satellite missions and for linking satellite location data to accurate geographical locations, such knowledge is essential. However, determining satellite orbits poses a significant challenge for small to medium satellite operators lacking the necessary tracking infrastructure. These organizations often rely on third-party services like Celestrak, which offer orbital information but may not provide it with the required frequency. Moreover, in the initial stages of a mission, especially when multiple satellites are launched together, it becomes challenging to attribute specific orbital parameters to individual satellites within tshe group. This ambiguity hampers mission planning and monitoring efforts. To address these challenges, this research presents an approach to tackle the problem of orbital parameter determination by leveraging Global Positioning System (GPS) data, and artificial intelligence, specifically genetic algorithms, to enhance the accuracy and efficiency of orbit determination algorithms and orbit propagation techniques, thus enabling small and medium satellite operators to reliably obtain vital orbital parameters.
Traditional methods for pattern recognition have demonstrated significant advancements in recent years. However, these techniques have traditionally relied on human intervention to discern crucial insights from data. Deep learning has ushered in a transformative shift by empowering computers to autonomously glean knowledge from data, particularly benefiting our understanding of how individuals interact with mobile and wearable technology. The burgeoning popularity of deep learning stems from its ability to operate effectively with minimal or no human guidance. This research introduces an innovative hybrid deep learning architecture that seamlessly integrates Convolutional Neural Network (CNN) and Multi-Layer Perceptron (MLP) layers. While maintaining precision in activity identification, this CNN-MLP approach excels in localized feature extraction, made possible through the synergy of CNN and MLP layers. Our model delivered outstanding performance on the UCI HAR dataset, achieving a classification accuracy of 97.14%.
Code-mix is gaining popularity due to its widespread usage on social media platforms. Even bilingual users converse in code-mix in their day-to-day life due to the popularity of the English language mix with the user’s mother tongue. Codemix phenomena come under NLP(Natural Language Processing), again a subpart of Artificial Intelligence. Code-mix language means grammatically following one language but has used another language’s words. Identification of code-mix is challenging due to using two languages in one sentence. Our motive is to identify the code-mix language at the sentence level. Already machine learning algorithms have been applied to identify sentence-level code-mix languages. In this article, we have applied Deep learning and a transformer-based framework for the sentence-level language identification task. We have used BERT in the Gujarati code-mix data set and received 96.2%, which is a more accurate result than the machine learning and Deep learning algorithms have achieved.
A deepfake is a piece of digital material that has been altered, such as a picture or a video in which the subject’s likeness has been substituted. Deepfake is a real danger to society since it distorts the opinions and perceptions of those around us. Deep learning, which is referred to as “a subset of AI,” is used to create deepfake. Deep learning is a combination of algorithms with the ability to learn and make sensible decisions on their own. People may be led astray by deepfake into believing something to be genuine when it is not. In the hands of internet scammers and cybercriminals, such cutting-edge technologies as deepfake can become dangerous tools. Deepfakes can be hard to spot, making it tough to stop the dissemination of fake content. Additionally, anyone with little programming knowledge can make deepfakes because they can be made with relatively simple tools. Understanding the dangers of deepfake, we are motivated to understand the technology behind deepfake generation, search for technologies to counter this threat, and develop effective detection techniques to accurately identify deepfake videos with high efficiency. In order to achieve this, we have conducted an extensive literature search on various algorithms for creating and detecting deepfake videos. We have identified their relative strengths and limitations, and areas for potential improvements. We have developed to enhance the current detection techniques by designing and developing a new technique that can effectively identify fake videos with multiple faces.
In pediatric rehabilitation, this paper introduces a novel approach to traditional Spider-cage therapy. This approach involves the integration of a proof-of-concept sensory system which would provide the therapists with quantitative measurements to assist them in directing the rehabilitation process rather than purely relying on their subjective judgments. This system incorporates an array of Force-Sensitive Resistor (FSR) sensors for weight distribution measurement and Galvanic Skin Response (GSR) sensors to measure pain levels. Through interactive visual and auditory stimulation, the patients are guided to execute a series of exercises seamlessly integrated within an interactive story-line. With a focus on patient-specific and evidence-based decision-making, the initial results show that this comprehensive approach promises to reshape paediatric rehabilitation, fostering adaptability, and minimizing subjectivity while enhancing overall therapeutic outcomes.
Recent advancements in image-based Facial emotion recognition systems offer deep insights into human psychological states, holding promise for valuable applications such as measuring customer satisfaction in public service areas or student engagement in classrooms. However, the potential impacts on the privacy of individual users of these systems cannot be ignored. In our study, we present PEEP (Privacy using EigEnface Perturbation) integrated with cloud storage encryption as a dual-layered approach to address privacy concerns in such systems. PEEP processes facial data by extracting eigenfaces and adding specific noise, ensuring the anonymity of identities while retaining the capability to recognize emotions. In tandem, cloud storage encryption guarantees that data, if intercepted during transmission or storage, stays encrypted and secure. This combined strategy offers an enhanced privacy solution for emotion recognition systems on remote servers. This study aligns with international initiatives to promote the responsible development and application of artificial intelligence, while emphasizing the importance of upholding human ethical standards and security.
Road sign detection and recognition play a critical role in improving driver safety and awareness in the modern traffic era. This paper describes the development and evaluation of a Road Sign Detection and Recognition System (RSDRS). Our system leverages computer vision techniques and mobile application technology to provide drivers with real-time visual and auditory feedback based on detected traffic signs. The implementation of RSDRS involves two basic phases: developing a robust recognition model and creating an Android application. The model, trained with the German Traffic Sign Recognition Benchmark (GTSRB) dataset and the YOLOv5s object detection algorithm, serves as the core component for accurate traffic sign recognition. The Android application captures live video frames, integrates seamlessly with the recognition model, and provides intuitive feedback to the driver. The performance evaluation reveals the exceptional capabilities of our system, with an average accuracy of 99.3% and a fast response time of 0.73 milliseconds. Metrics such as recall, precision, and F1 score highlight the model’s ability to maintain accuracy while minimizing false positives. Real-world applicability is paramount, and our system excels in a variety of environments and lighting conditions, as evidenced by rigorous testing.
Recommender systems (RSs) are one of the most important tools for helping users decide what to buy, play, read, book, and watch. They become very effective tools for information filtering. The aim of this work is to forecast user ratings for a variety of movies, which is a current research topic in collaborative filtering (CF). Numerous CF models have been analysed and modeled in this paper for empirical-based comparison analysis. We have investigated several similarity metrics based on user-item rating predictions. We will then compare them by running simulations on the MovieLens dataset and we will use the root mean squared errors (RMSE), the precision, and the Spearman’s rank correlation coefficient as comparison statistics. A perfect correlation and RMSE of 0 were achieved through experimental analysis. A further advantage of the model is that it predicted the top 50 movie ratings with a precision of 100%.
Edge computing provides benefits such as reduced latency by processing data closer to its source, but it faces challenges due to limited processing power, memory, and storage capacity. To address these limitations and enhance system efficiency, task classification becomes crucial in edge computing. By categorizing tasks, resources can be allocated effectively. This paper presents a novel approach for classifying tasks and edge servers based on CPU and memory capacities. The proposed model treats tasks and edge servers as points and applies the K-means method by adding multi-objective constraints Chebyshev distance. The distance metric considers two objectives: the distance between points and the number of tasks and edge servers. Unlike traditional K-means algorithms that often result in clusters with either one server or clusters without servers, our model ensures that nearly every cluster contains at least one server. Simulation results demonstrate the fast convergence of the proposed model, evaluated through clustering using inertia and silhouette coefficient. It is important to note that this classification does not directly influence task scheduling or resource allocation processes. Instead, it serves as a preliminary step to improve the effectiveness of these subsequent processes. By strategically placing more powerful tasks on higher-capacity servers, and vice versa, our approach aims to reduce the workload of the scheduler and enhance resource allocation or task scheduling. This classification framework has the potential to achieve efficient task and resource management in edge computing environments.
In this study, we examine different transformer based pretrained Artificial Intelligence (AI) models on their ability to summarize text content from different sources. AI has emerged as a powerful tool in this context, offering the potential to automate and improve the process of content summarization. We mainly focus on the pretrained transformer models, such as Pegasus, T5, Bart, and ProphetNet for key point summarization from textual contents. We aim to assess the effectiveness of these models in summarizing different contents like articles, instructions, conversational dialogues, and compare and analyze their performance across different datasets. We use ROUGE metric to evaluate the quality of the generated summaries. The Facebook’s BART model had better performance across different textual datasets. We believe that our findings will offer valuable insights into the capabilities and limitations of Transformer-based AI models in the context of extracting essential points from large articles, making them useful as assistive tools for summarizing course content in educational environments.
Misinformation, also known as "fake news," is a growing problem in social media. It refers to the dissemination of false information, intentionally or unintentionally, in digital media platforms such as Facebook, Twitter, and Instagram. The issue has gained global attention due to its potential to cause harm and pose a threat to democracy. Misinformation in social media has significant consequences such as loss of trust in institutions and the stoking of fear and anger in people. The causes and mechanisms of spreading fake news are complex, involving psychological, social, and technological factors. To tackle the problem, it is necessary to develop interventions that address these factors and to enhance media literacy among the public, so people can tell fact from fiction. The objective of this research paper is to perform a comprehensive review of the techniques and tools that are used to cause, spread, and detect misinformation in social media, as well as perform in-depth experiments with the open-source tools and datasets to evaluate and compare those techniques. We believe the research work’s outcome will benefit researchers, educators, and students through dissemination of the knowledge and resources, including the open-source tools and data acquired by the research work.
Steganography, the practice of concealing secret information within innocuous carriers, exploits inherent data redundancies to ensure effective covert communication. However, challenges like lossy compression and data loss can compromise the integrity of the hidden data and complicate accurate recovery. Thus, this paper introduces an new technique in order to retrieve lost data. This approach introduces a novel sender-receiver communication mechanism to extract secret images from steganographic media while compensating for lost portions. The proposed technique enables precise retrieval of confidential images even in scenarios involving significant data loss. Extensive experimental evaluation conducted across diverse steganographic images substantiate the approach’s effectiveness in minimizing distortion during secret image recovery. Notably, the method yields a substantial improvement in the Peak Signal-to-Noise Ratio (PSNR), attesting to enhanced image quality and likeness to the original covert image. Moreover, the collaborative sender-receiver process bolsters recovery accuracy, curbing data loss and preserving the secret image’s integrity.
This research project proposes the use of interactive videos to enhance teachers' training in educating autistic children. Effective teaching methods for students with autism require teachers to comprehend the condition and employ tailored instructional strategies, including adapting assignments, aiding those with language difficulties, and utilizing visual aids for better organization and focus. Traditional teacher training methods can be both expensive and time-consuming. In contrast, interactive videos provide a proactive and flexible way to access training content, empowering teachers to engage with the material dynamically and take control of their learning experiences. Future work will explore the integration of AI-driven ChatGPT to offer personalized support and create a dynamic training program, with the goal of improving inclusivity and educational quality in autism settings while benefiting teachers, students, and the education system at large.
This research paper investigates the security principles of a chatroom application. The study aims to ensure the security and privacy of user information while maintaining user convenience and ease of use. A literature review and analysis identified potential security vulnerabilities, including unauthorized access to user accounts, insecure transmission and storage of messages, and possible man-in-the-middle attacks or data breaches. To address these challenges, a new secure chatroom application is proposed. It integrates an in-depth security strategy to ensure high communication protection. It includes a) Deployment of user authentication and password security level verification, b) Integration of encryption to ensure secure messaging and user data storage using Advanced Encryption Standard and SHA-256 hashing, respectively, c) Adoption of no message history features to ensure the availability of only current conversation and no older message are retrieved, d) Integration of client handler to manage the interaction between client and server and ensure efficient and smooth chatroom operations, and d) a user-friendly GUI that serve the basic of communication with high-security level and low resource consumption. The research findings suggest that a balance can be achieved between security and user convenience, resulting in a chatroom application that is both secure and easy to use.
Epilepsy is a prevalent neurological disorder and has been studied through the analysis of Electroencephalogram (EEG) signals. However, the identification and classification of epileptic seizure patterns remains challenging due to the non-stationary nature of EEG signals and the presence of artifacts. In this paper, we investigate the applicability of a transformer-based deep learning model to classify seizure patterns observed in epileptic patients. We employed the self-attention mechanism inherent in transformers to capture complex temporal relationships in the EEG recordings. By prepossessing the EEG signals into suitable input sequences and adapting the transformer architecture, we achieved 78.11% in distinguishing between different epileptic seizure patterns. Our findings indicate that the transformer model, with its ability to manage long-range dependencies, offers a robust approach to EEG-based seizure pattern classification. This work is important for building advanced automated diagnostic tools for epilepsy and related neurological disorders.
