IAES International Journal of Artificial Intelligence (IJ-AI)
Not a member yet
1769 research outputs found
Sort by
Video forgery: An extensive analysis of inter-and intra-frame manipulation alongside state-of-the-art comparisons
Get PDFThe widespread accessibility of inexpensive mobile phones, digital cameras, camcorders, and security closed-circuit television (CCTV) cameras has resulted in the integration of filmmaking into our everyday existence. YouTube, Facebook, Instagram, and Snapchat are a few of the video-sharing and editing applications that facilitate the process of uploading and editing videos. Additional instances include Adobe Photoshop, Windows Movie Maker, and Video Editor. Although editing has its advantages, there is a potential risk of counterfeiting. This occurs when films are edited with the intention of misleading viewers or manipulating their perspectives, which can be particularly troublesome in judicial procedures where recordings are submitted as evidence. The issue has been exacerbated by the emergence of deep learning methods, such as deepfake videos that effectively manipulate facial characteristics. Consequently, individuals have become less reliant on visual evidence. These issues emphasise the pressing necessity for the creation of dependable methods to determine the authenticity of films and identify cases of fraud. Contemporary methods can depend on assessing modified frames or utilising distortions generated during video codec compression or double compression. Since 2016, multiple studies have been undertaken to investigate techniques, strategies, and applications to tackle this problem. The objective of this survey study is to provide a comprehensive analysis of these algorithms, highlighting their advantages and disadvantages in detecting different forms of video forgeries
An enhanced cascade ensemble method for big data analysis
Get PDFIn the digital age, the proliferation of data presents both challenges and opportunities, particularly in the realm of big data, which is characterized by its volume, velocity, and variety. Machine learning is a crucial technology for extracting insights from these vast datasets. Among machine learning methods, ensemble methods, and especially cascading ensembles, are highly effective for big data analysis. While it is true that the training procedures for cascade ensembles can be time-consuming and may have limitations in terms of accuracy, this paper proposes a solution to enhance their performance. Our method involves using stochastic gradient descent (SGD) classifiers, an improved training data separation algorithm, and integrating principal component analysis (PCA) at each ensemble level. We are confident that these enhancements lead to improved results and accuracy. The proposed approach is designed to enhance both the generalization properties and accuracy of the ensemble (3%), while also reducing its training time. Results from modelling on a real-world biomedical dataset demonstrate significant reductions in training duration, improvements in generalization properties, and enhanced accuracy when compared to other possible implementations of the ensemble
Serial parallel dataflow-pipelined processing architecture based accelerator for 2D transform-quantization in video coder and decoder
Get PDFThe video coder and decoder (CODEC) standards from MPEG-4 to the recent versatile video codec (VVC), adopted lossy compression methodologies, which involves transformation, quantization and entropy coding. The growing usage of video data in all means of communication demands more bandwidth and storage requirements. While compression with redundancy removal by transform coefficient coding, the focal point is the crucial sequential data flow and data processing structures. Handling the block wise data near to the processing unit prior and after computation will reduce the data waiting time of the processing unit, hence accelerating the targeted functionality. The proposed serial parallel data-flow pipelined processing architecture (SPDPA) accelerates the speed of processing unit by on chip data availability and parallel data accessing options and also with the pipeline operations of transformation, data transpose and quantization. The post implementation results of the architecture targeted to 16 nm and 28 nm field programmable gate array (FPGA) shows that there is a trade-off between power and frequency of operations for various block sizes. The design targeted to 16 nm works for higher frequencies with an average power consumption 0.64 w as compared to 28 nm FPGA which consumes less average power of 0.15 w
Semi-automatic voice comparison approach using spiking neural network for forensics
Get PDFThis paper explores the application of a semi-automatic technique using spiking neural network (SNN) approach for forensic voice comparison (FVC), addressing the limitations of traditional methods that are time-consuming and subjective. By integrating machine learning with human expertise, the SNN, which mimics the brain’s processing of temporal information, is applied to analyze Australian English voice data in .flac format. The model leverages synaptic connection strengths modified by spike timing, allowing for flexible voice feature representation. Performance metrics, including confusion matrices and receiver operating characteristic (ROC) analysis, indicate the model’s accuracy of 94.21%, highlighting the effectiveness of the SNN-based approach for FVC
A comparative study of large language models with chain-of thought prompting for automated program repair
Get PDFAutomatic code repair is an important task in software development to reduce bugs efficiently. This research focuses on developing and evaluating a chain-of-thought (CoT) prompting approach to improve the ability of large language models (LLMs) in automated program repair (APR) tasks. CoT prompting is a technique that guides LLM to generate step-by-step explanations before providing the final answer, so it is expected to improve the accuracy and quality of code repair. This research uses the QuixBugs dataset to evaluate the performance of several LLM models, including DeepSeek-V3 and GPT-4o, with two prompting methods, namely standard and CoT prompting. The evaluation is based on the average number of plausible patches generated as well as the estimated token usage cost. The results show that CoT prompting improves performance in most models compared with the standard. DeepSeek-V3 recorded the highest performance with an average of 36.6 plausible patches and the lowest cost of 0.226. These results confirm that CoT prompting is an effective technique to improve LLM reasoning ability in APR tasks
Prediction of flood-affected areas based on geographic information system data using machine learning
Get PDFFlood disasters have become more frequent and severe due to climate variability, posing significant threats to human lives, agriculture, and infrastructure. Effective disaster management and mitigation require accurate identification of flood-prone areas. This study develops an intelligent flood prediction system by integrating machine learning algorithms with geographic information systems (GIS) data to enhance flood risk assessment. The proposed system utilizes two machine learning models, including random forest (RF) and support vector machine (SVM), to predict flood-susceptible areas. The models are trained on historical flood data and GIS-derived features, including elevation, slope, topographic wetness index (TWI), aspect, and curvature. The dataset undergoes preprocessing, including normalization and feature selection, before being divided into training, validation, and test sets. The models are then trained and evaluated based on their predictive performance. Evaluation metrics, particularly the area under the curve (AUC), demonstrate that RF outperforms SVM in predicting flood-prone areas. RF achieves an accuracy of 82%, while SVM records a lower accuracy of 68%. The superior performance of RF is attributed to its ability to handle complex, nonlinear relationships in flood prediction. These results highlight the effectiveness of machine learning algorithms in flood susceptibility modeling and support the integration of data-driven techniques into flood and disaster risk reduction management strategies
Deep learning-aided polar-low density parity check decoding for enhanced telemedicine ECG transmission reliability
Get PDFTelemedicine has emerged as a crucial solution for remote patient monitoring and diagnosis, yet ensuring the reliable transmission of medical data, particularly electrocardiogram (ECG) signals, remains a significant challenge. This work proposes a novel approach that integrates deep learning with a polar-low density parity check (LDPC) decoder to enhance the accuracy, robustness, and efficiency of ECG signal transmission within telemedicine systems. The study aims to evaluate the effectiveness of this integration in improving error correction and decoding performance, validate its efficacy under diverse signal to noise ratios (SNRs) and code rates, and assess its potential impact on remote healthcare delivery. Experimental results confirm that the deep learning-empowered polar-LDPC decoder achieves superior error correction and decoding efficiency compared to conventional methods, ensuring higher fidelity in ECG reconstruction. This advancement presents a promising pathway toward more reliable, precise, and efficient telemedicine systems, thereby enabling improved patient care, especially in remote and underserved regions. The proposed method also opens opportunities for integrating intelligent decision-support tools. Such integration could further enhance real-time diagnostics and broaden telemedicine’s scope
Solving sparsity and scalability problems for book recommendations on e-commerce
Get PDFThis study proposed a hierarchical density-based spatial clustering of applications with noise (HDBSCAN) and randomized singular value decomposition (RSVD) collaborative filtering (CF) method to overcome sparsity and scalability problems for book recommendations on e-commerce. CF is an information retrieval system that assumes a user has the same interest in an object as other users have in the past. When handling large volumes of data, sparsity problems can arise, where finding a similarity relation of user preferences results from a small assessment of an object by users. The scalability is the increased computation of an algorithm caused by increased users or objects, which makes recommendations take longer to form, therefore making them less accurate. HDBSCAN is a density-based clustering method that simplifies the hierarchical arrangement of the most significant clusters for extraction to group users in the same cluster. RSVD is a linear dimension reduction method that breaks a matrix into three sub matrices by reconstructing the size of that matrix without removing its dominant part, especially for cluster result matrices. The HDBSCAN RSVD-CF model reduced the root mean squared error (RMSE) by 21.83%, being 3793.73 seconds faster than the CF model. It also performed very well compared to both RSVD-CF and HDBSCAN-CF
A review of driver distraction detection while driving based on convolutional neural networks
Get PDFDriver distraction represents a major cause of traffic accidents, posing a serious threat to human life. In this review, we present the latest research findings of driver distraction detection based on convolutional neural networks (CNNs). In general, the analysis of driver behavior while driving is represented by either detecting driver drowsiness or attention diversion from driving by other activities, all of which fall under the definition of driver distraction. Facial features are often the basis for detecting driver drowsiness. In most papers, it is typically done by eye blinking, yawning, and head movement. As for the driver attention diversion, it is through the position of the hand and face. It involves many activities, text messages, making phone calls, adjusting the radio, consuming beverages, reaching for objects behind the driver, applying makeup, interacting with passengers, and other similar distractions. However, suggesting new methodologies in driver distraction detection and choosing appropriate CNN-based techniques is a big challenge given the wide variety experiments and studies in this field. Therefore, previous papers should be revisited to produce new methods by taking advantage of the techniques used. As a result, this paper reviews research approaches and reveals the effectiveness of CNN in detecting driver distraction. Finally, the article lists techniques that can be used as benchmarks in this context
Hybrid N-gram-based framework for payload distributed denial of service detection and classification
Get PDFThere are three primary approaches to DDoS detection: anomaly-based, pattern-based, and heuristic-based. The heuristic-based method integrates both anomaly- and pattern-based techniques. However, existing DDoS detection systems face challenges in performing HTTP payload-level analysis, mainly due to high false positive rates and insufficient granularity in current datasets. To address this, the study introduces a novel heuristic approach based on a hybrid N-Gram model. This hybrid combines two components: CSDPayload+N-Gram and CSPayload+N-Gram. CSDPayload represents the gap (measured via Chi-Square Distance) between a given payload and normal traffic payloads, while CSPayload reflects the similarity (measured via Cosine Similarity) between them. These metrics form a new feature set evaluated using three datasets: CIC2019, MIB2016, and H2N-Payload. The methodology begins with packet extraction and conversion of TCP/IP traffic—specifically HTTP traffic—into hexadecimal payloads. N-Gram analysis (from 1-Gram to 6-Gram) is then applied to these payloads. For each N-Gram, frequency counts are computed, followed by calculations of Chi-Square Distance (CSD), Cosine Similarity (CS), and Pearson’s Chi-Square test to classify payloads as either benign or malicious. Subsequently, feature selection is performed using weight correlation, and the resulting features are fed into three machine learning classifiers: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Neural Network. Experimental results demonstrate high detection accuracy, particularly in the 4-Gram feature category: Neural Network achieves 99.65%, KNN 95.14%, and SVM 99.73% accuracy on average