Abstract: The increase in malicious cyber activities has generated the need to produce effective tools for the field of digital forensics and incident response. Artificial intelligence (AI) and its fields, specifi- cally machine learning (ML) and deep learning (DL), have shown great potential to aid the task of processing and analyzing large amounts of information. However, models generated by DL are often considered “black boxes”, a name derived due to the difficulties faced by users when trying to understand the decision-making process for obtaining results. This research seeks to address the challenges of transparency, explainability and reliability posed by black box models in digital forensics. To do this, explainable artificial intelligence (XAI) is explored as a solution. This approach seeks to make DL models more interpretable and understandable by humans. The SHAP (SHapley Additive eXplanations) and LIME (Local Interpretable Model-agnostic Explanations) methods will be implemented and evaluated as a model-agnostic technique to explain predictions of the generated models for forensic analysis. In order to make transparent the decision-making process in the models, and to evaluate the confidence of the generated results

Abstract: Self-similar data streams are characterized by their similarity across multiple time scales, exhibiting distinct nonlinear and discrete features. These characteristics complicate the accurate identification of data points associated with soft error features, thereby making it difficult to effectively discern the intricate relationship between the data flow and soft error data. This, in turn, severely impacts the accuracy of soft error detection in self-similar data streams. In this study, we propose a novel approach to address this challenge. Leveraging the suddenness and long-range correlation inherent in self-similar data streams, we construct a time series model to capture the data stream based on linear correlation and straight-line fitting features. By incorporating relationship parameters to fit neighboring flow points, we utilize a structural mapping model to establish the local angular relationship between the data streams and soft error data. Additionally, we construct a structural mapping network using flow features to achieve soft error detection in self-similar data streams. Experimental results demonstrate that our proposed method achieves high accuracy and low time overhead for soft error detection in self-similar data streams.

Abstract: China is the world's largest producer of chili peppers, which occupy particularly important economic and social values in various fields such as medicine, food and industry.However, during its production process, chili peppers are affected by pests and diseases resulting in significant yield reduction due to temperature, environment and other reasons. In this study, a lightweight pepper disease identification method DD-YOLO based on the YOLOv8n model is proposed. First, the deformable convolutional module DCNv2 (Deformable Convolutional Networks) and the inverted residual mobile block iRMB (Inverted Residual Mobile Block) are introduced into the C2Fmodule to improve the accuracy of the sampling range and reduce the computational amount; secondly, the DySample sampling operator (Dynamic Sample) is integrated into the head network toreduce the amount of data and reduce the complexity of computation. Finally, we use Large Separable Kernel Attention (LSKA) to improve the SPPF module (Spatial Pyramid Pooling Fast) to enhance the performance of multi-scale feature fusion. The experimental results show that the accuracy, recall and average precision of the DD-YOLO model are 91.6%, 88.9% and 94.4%, respectively, compared with the base network YOLOv8n, it improves 6.2, 2.3 and 2.8 percentage points respectively, the model weight is reduced by 22.6%, and the number of floating-point operations per second is improved by 11.1%. This method provides a technical basis for intensive cultivation and management of chili peppers as well as efficiently and cost-effectively accomplishes the task of identifying chili pepper pests and diseases.

Abstract: Electronic waste (e-waste) is one of the fastest-growing waste streams worldwide, posing critical environmental, economic, and public health challenges. The circular economy paradigm offers a holistic approach to managing e-waste through resource recovery, recycling, and reduced landfill disposal. Recently, Artificial Intelligence (AI) and Machine Learning (ML) have demonstrated transformative potential in addressing central bottlenecks in e-waste handling, including precise materials identification, automated disassembly, improved recycling efficiency, and predictive logistics. This paper critically evaluates 30 peer-reviewed studies published between 2010 and 2025, selected via a transparent screening process, focusing on AI- and ML-driven technologies for e-waste management within the circular economy. We synthesize evidence from real-world implementations, discuss performance metrics (e.g., sorting accuracy, throughput gains, and carbon footprint reduction), and highlight how AI and ML algorithms can boost recovery of high-value materials, reduce environmental impact, and improve overall cost-effectiveness. We further examine current trends, underscore notable achievements, and analyze key challenges—such as data privacy, regulatory gaps, heterogeneous waste streams, and algorithmic bias. A series of policy recommendations and a future research roadmap are proposed, delineating technological, regulatory, and socio-economic pathways to expedite adoption of AI-enhanced e-waste management. By presenting a rigorous, thematically focused synthesis, this review anchors AI-based e-waste solutions as a linchpin for advancing the circular economy and achieving sustainable development.

Abstract: Customer churn poses a significant challenge across various sectors, resulting in considerable revenue losses and increased customer acquisition costs. Machine Learning (ML) and Deep Learning (DL) have emerged as transformative approaches in churn prediction, significantly outperforming traditional statistical methods by effectively analyzing high-dimensional and dynamic customer datasets. This literature review systematically examines recent advancements in churn prediction methodologies based on 240 peer-reviewed studies published between 2020 and 2024 across diverse domains such as telecommunications, retail, banking, healthcare, education, and insurance. It emphasizes the evolution of ML and DL approaches, their practical applications, and ongoing challenges, including model interpretability, class imbalance, and concept drift. The study identifies an increasing preference for advanced techniques, including ensemble models, profit-driven frameworks, hybrid architectures, and sophisticated DL methods like convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and attention mechanisms. It highlights the growing focus on explainable AI (XAI), profit-oriented modeling, and adaptive learning strategies to accommodate evolving customer behaviors. Despite these advancements, the review underscores persistent challenges such as class imbalance, the black-box nature of complex DL models, difficulties adapting to concept drift, and limited consideration of real-world deployment constraints. This review contributes to the field by comprehensively synthesizing recent methodological trends and identifying gaps related to real-world applicability, interpretability, and business-oriented evaluation metrics. It offers essential insights and practical guidance for data scientists, researchers, and industry practitioners seeking to develop more accurate, robust, and interpretable churn prediction models, enabling more effective customer retention strategies and improved business outcomes.

Abstract: Generating new video games using AI has potential to be the next holy grail of the video game industry. Current AI efforts have focused on two directions: i) controllable video generation and ii) code generated by Large Language Models (LLMs). The first direction is limited due to short-term memory and increasing corruption (blur, noise) over time. The second direction is promising, but it requires a lot of human hand-holding with human-provided assets. Generating hours of coherent interactive video content is infeasible. In this paper, we instead attempt the overly ambitious problem of end-to-end generation of Small Web Format (SWF) games and animations through bytes. By modeling bytes, one does not need code or assets to potentially obtain full games with title screen, narrative, text, graphics, music, and sounds. We make a first attempt by fine-tuning a 7-billion-parameter LLM at 32K context length to generate the bytes of video games and animations conditional on a text description. Our model (ByteCraft) can generate up to 32K tokens, each containing at most 4-5 bytes (generating files as big as 140 KB). Some of the generated files are partially working (4.8-12%), or fully working (0.4-1.2%). ByteCraft is a proof-of-concept highlighting what could be possible given more scaling and engineering effort. We open-source our model and inference code alongside a dataset of 10K synthetic prompts for use with ByteCraft.

Abstract: Developing reliable AI systems to assist human clinicians in multi-modal medical diagnosis has long been a key objective for researchers. Recently, Multi-modal Large Language Models (MLLMs) have gained significant attention and achieved success across various domains. With strong reasoning capabilities and the ability to perform diverse tasks based on user instructions, they hold great potential for enhancing medical diagnosis. However, directly applying MLLMs to the medical domain still presents challenges. They lack detailed perception of visual inputs, limiting their ability to perform quantitative image analysis, which is crucial for medical diagnostics. Additionally, MLLMs often exhibit hallucinations and inconsistencies in reasoning, whereas clinical diagnoses must adhere strictly to established criteria. To address these challenges, we propose MedAgent-Pro, an evidence-based reasoning agentic system designed to achieve reliable, explainable, and precise medical diagnoses. This is accomplished through a hierarchical workflow: at the task level, knowledge-based reasoning generate reliable diagnostic plans for specific diseases following retrieved clinical criteria. While at the case level, multiple tool agents process multi-modal inputs, analyze different indicators according to the plan, and provide a final diagnosis based on both quantitative and qualitative evidence. Comprehensive experiments on both 2D and 3D medical diagnosis tasks demonstrate the superiority and effectiveness of MedAgent-Pro, while case studies further highlight its reliability and interpretability. The code is available at https://github.com/jinlab-imvr/MedAgent-Pro.

Abstract: Identifying diseases in horticulture fruits is crucial in maintaining quality, reducing losses, and enhancing sustainable agricultural practices. Deep learning (DL) and machine learning (ML) techniques have enabled proficient and precise identification of these diseases. This paper consolidates the use of ML and DL approaches in horticultural fruit disease detection, incorporating the innovative models of convolutional neural networks (CNNs), Vision transformers, and other hybrid systems. It also reviews preprocessing and feature extraction for hyperspectral and multispectral imaging. Volume public datasets and real-world case studies are analyzed to demonstrate practical implementation and obstacles which include the quality of the dataset, required computation resources, and model interpretability. Furthermore, the paper elaborates on GAN-based data augmentation, implementing lightweight models on resource-constrained devices, and real-time IoT monitoring. Future directions aim at the utilization of explainable artificial intelligence, scaling up the models, and increasing sustainability in disease detection systems. The reviewed literature established this study as a point of reference for other researchers and practitioners to inspire the development of intelligent horticultural disease management systems.

Abstract: Explainability is increasingly crucial for real-world deployment of deep learning models, yet traditional explanation techniques can be prohibitively slow and memory- intensive on resource-constrained devices. This paper presents a novel lightweight ex- plainability framework that significantly reduces the computational cost of generating explanations without compromising on quality. My approach focuses on an optimized Grad-CAM pipeline with sophisticated thresholding, advanced memory handling, and specialized evaluation metrics. I demonstrate speedups exceeding 300x over naive im- plementations while maintaining robust faithfulness and completeness scores. Through an extensive series of benchmarks, user studies, and statistical tests, I show that this framework is scalable, accurate, and deployable on edge devices such as Raspberry Pi, Android phones, and iPhones. I also discuss ethical considerations, future research directions, and potential applications in high-stakes domains like healthcare and au- tonomous systems.

Abstract: The proliferation of AI tools in academic writing poses significant challenges for verifying the authenticity of student submissions, particularly at the sentence level. This paper proposes a novel Stylometric Fingerprinting with Contextual Anomaly Detection approach to distinguish AI-generated sentences from student-authored ones in writing reports. By combining manual stylometric analysis with contextual coherence checks, our method achieves sentence-level granularity without requiring computational tools or LLMs, making it accessible for student use. We compare our approach to existing models like Turnitin, GPTZero, and Moss, highlighting its unique focus on manual, coherence-driven detection. Experimental insights and theoretical analysis demonstrate its feasibility and effectiveness in ensuring academic integrity.

Abstract: This paper provides an extensive review of DeepSeek, an emerging open-source large language model (LLM) known for its Mixture-of-Experts (MoE) architecture and Multi-Head Latent Attention innovations. The study highlights DeepSeek's superior efficiency, scalability, and performance across tasks such as natural language processing, mathematical reasoning, and code generation, positioning it as a competitive alternative to proprietary models like ChatGPT, Claude, and Gemini. Comparative evaluations reveal its strengths in formal writing, structured reasoning, and diagnostic applications in healthcare and finance, while noting challenges in creative tasks and user safety concerns. With a focus on democratizing AI, DeepSeek's cost-efficient, open-source nature fosters accessibility and collaboration across industries such as education, business, and healthcare. Ethical considerations and future directions, including multimodal integrations and enhanced safety protocols, are also explored. Overall, the paper underscores DeepSeek's potential in driving innovation and expanding the frontiers of artificial intelligence research and applications. Comparative analyses reveal that DeepSeek excels in tasks requiring structured writing, grammatical precision, and technical problem-solving. For instance, it achieves notable success in healthcare diagnostics and risk management in finance. However, challenges include its limitations in creative outputs and a higher rate of unsafe responses compared to some competitors, signaling the need for enhanced safety protocols. The paper also highlights user feedback, which is generally positive regarding accessibility and reasoning capabilities, though criticisms are directed at content policies and moderation. DeepSeek's open-source nature is celebrated for democratizing AI, making advanced technology accessible to researchers, educators, and developers worldwide, particularly in resource-constrained settings. Applications across education, healthcare, and finance demonstrate its versatility, from personalizing learning experiences to improving diagnostic accuracy and enabling better financial decision-making. Future directions include expanding its multimodal capabilities, refining safety measures, and exploring innovative applications to maximize its impact across industries.

Abstract: Personalized recommendation systems play a crucial role in enhancing user engagement and decision-making across various domains. Traditional approaches, such as collaborative filtering and matrix factorization, have shown effectiveness but suffer from data sparsity and cold-start problems. Recent advances in deep learning, graph-based models, and attention mechanisms have significantly improved recommendation performance. This paper proposes a novel hybrid recommendation model that integrates Factorization Machines (FM), Graph Convolutional Networks (GCN), and Multi-Layer Attention Networks (MLAN) to optimize feature representations and enhance prediction accuracy. Experimental results demonstrate the superiority of the proposed approach over baseline methods in key performance metrics.

Abstract: Emotion recognition in speech is essential for enhancing human-computer interaction (HCI) systems. Despite progress in Bangla speech emotion recognition, challenges remain, including low accuracy, speaker dependency, and poor generalization across emotional expressions. Previous approaches often rely on traditional machine learning or basic deep learning models, struggling with robustness and accuracy in noisy or varied data. In this study, we propose a novel multi-stream deep learning feature fusion approach for Bangla speech emotion recognition, addressing the limitations of existing methods. Our approach begins with various data augmentation techniques applied to the training dataset, enhancing the model’s robustness and generalization. We then extract a comprehensive set of handcrafted features, including Zero-Crossing Rate (ZCR), chromagram, spectral centroid, spectral roll-off, spectral contrast, spectral flatness, Mel-Frequency Cepstral Coefficients (MFCCs), Root Mean Square (RMS) energy, and Mel-spectrogram. These features capture key characteristics of the speech signal, providing valuable insights into the emotional content. Sequentially, we utilize a multi-stream deep learning architecture to automatically learn complex, hierarchical representations of the speech signal. This architecture consists of three distinct streams: the first stream uses 1D Convolutional Neural Networks (1D CNN), the second integrates 1D CNN with Long Short-Term Memory (LSTM), and the third combines 1D CNN with Bidirectional LSTM (Bi-LSTM). These models capture intricate emotional nuances that handcrafted features alone may not fully represent. For each of these models, we generate predicted scores, and then employ ensemble learning with a soft voting technique to produce the final prediction. This fusion of handcrafted features, deep learning-derived features, and ensemble voting enhances the accuracy and robustness of emotion identification across multiple datasets. Our method demonstrates the effectiveness of combining various learning models to improve emotion recognition in Bangla speech, providing a more comprehensive solution compared to existing methods. We utilize three primary datasets—SUBESCO, BanglaSER, and a merged version of both—as well as two external datasets, RAVDESS and EMODB, to assess the performance of our models. Our method achieves impressive results with accuracies of 92.90%, 85.20%, 90.63%, 67.71%, and 69.25% for the SUBESCO, BanglaSER, merged SUBESCO and BanglaSER, RAVDESS, and EMODB datasets, respectively. These results demonstrate the effectiveness of combining handcrafted features with deep learning-based features through ensemble learning for robust emotion recognition in Bangla speech.

Abstract: This study investigates the primary technological and socio-environmental factors influencing the adoption intentions of AI-powered technology at the corporate level within higher education institutions. A conceptual model based on the Diffusion of Innovation Theory (DOI) and the Technology-Organization-Environment (TOE) framework was proposed and tested using data collected from 367 higher education students, faculty members, and employees. The findings reveal that Compatibility, Complexity, User Interface, Perceived Ease of Use, User Satisfaction, Performance Expectation, AI introducing new tools, AI Strategic development, Availability of Resources, Technological Support, and Facilitating Conditions significantly impact AI adoption intentions. At the same time, Competitive Pressure and Government Regulations do not. Demographic factors, including major and years of experience, moderated these associations, and there were large differences across educational backgrounds and experience. The SPSS Amos 24 was used for SEM to choose the best-fitting model that proved to be more efficient than traditional multiple regression analysis.

Abstract: This study presented a comprehensive investigation into prompt engineering for large language models (LLMs) within educational contexts, combining a systematic literature review with a 12-week empirical study involving primary school students using a chatbot-based tutor in a Python programming course. The research explored the breadth of prompt engineering techniques, identified essential components for effective educational prompts, examined strategic applications, highlighted key implementation challenges, and captured learner perspectives on interacting with LLMs.Our review categorized prompt engineering techniques into foundational (e.g., zero-shot, few-shot, and direct instruction), structured reasoning (e.g., chain-of-thought, tree-of-thought, and graph-based models), hallucination reduction (e.g., retrieval-augmented generation, CoVe, ReAct), user-centric strategies (e.g., automatic prompt engineering, active prompting), and domain-specific applications (e.g., emotion prompting, contrastive reasoning, and code generation tools like PoT and CoC). We also examined advanced optimization methods including prompt tuning, abstraction, and self-consistency approaches that enhanced both reasoning and factual reliability.Key components of effective educational prompt engineering were distilled into nine categories: content knowledge, critical thinking, iterative refinement, clarity, creativity, collaboration, digital literacy, ethical reasoning, and contextual integration. These elements collectively supported both the quality of LLM outputs and the development of students’ cognitive and metacognitive skills.Strategically, we identified ten educational prompt engineering practices—contextual framing, task segmentation, prompt sequencing, role-based prompting, reflection, counterfactual exploration, constraint-based creativity, ethical consideration, interactive refinement, and comparative analysis—as essential for guiding LLM interactions aligned with pedagogical goals.We also addressed core challenges in prompt engineering for education, including ambiguity in model interpretation, balancing specificity and flexibility, ensuring consistency, mitigating hallucinations, safeguarding ethics and privacy, and maintaining student engagement. These challenges highlighted the need for explicit instructional support and adaptive prompt design in classrooms.Empirically, our study of primary school learners revealed a surprising level of sophistication in students’ prompt construction and refinement. Students developed intuitive understandings of prompt clarity, used context to guide AI responses, adopted role-based and scenario-based prompting, applied constraints to improve learning outcomes, and created reusable prompt templates. Furthermore, they engaged in iterative refinement, developed evaluation criteria for AI responses, and differentiated between general and specific prompts based on their learning objectives. These findings underscored students’ emerging metacognitive awareness and adaptability in AI-mediated learning.

Abstract: Distributed Denial of Service (DDoS) attacks pose a significant threat to cloud networks, disrupting services, degrading performance, and causing financial losses. Traditional security mechanisms, such as rule-based intrusion detection systems and traffic filtering, struggle to effectively counter evolving DDoS tactics due to the increasing complexity and volume of attacks. Artificial Intelligence (AI)-driven approaches provide a more adaptive and intelligent solution by leveraging machine learning (ML) and deep learning (DL) techniques to detect, classify, and mitigate DDoS attacks in real time.AI models analyze vast amounts of network traffic data to identify anomalous patterns, distinguishing between legitimate and malicious requests with high accuracy. Supervised and unsupervised learning techniques, such as Support Vector Machines (SVM), Random Forest, K-Nearest Neighbors (KNN), and neural networks, enhance threat detection by recognizing attack signatures and detecting zero-day attacks. Advanced deep learning architectures, including Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), further improve the accuracy and speed of anomaly detection by learning from historical attack patterns.In addition to detection, AI-powered mitigation strategies enable dynamic resource allocation, traffic rate limiting, and network traffic redirection. Techniques such as Reinforcement Learning (RL) and Federated Learning (FL) allow adaptive defense mechanisms that continuously evolve based on emerging attack vectors. Software-Defined Networking (SDN) and Network Function Virtualization (NFV) enhance the flexibility of AI-based solutions by enabling real-time monitoring and automated response mechanisms, minimizing downtime and ensuring service availability.This study explores the effectiveness of AI in fortifying cloud networks against DDoS attacks by integrating AI-driven threat intelligence with cloud security frameworks. By automating the detection and response process, AI enhances the scalability, efficiency, and resilience of cloud infrastructures, reducing the impact of DDoS attacks and ensuring uninterrupted service delivery. The integration of AI with cloud security measures represents a transformative approach to mitigating cyber threats in modern digital environments.

Abstract: Arabic poetry follows intricate rhythmic patterns called ‘arūḍ’ (prosody), so its automated categorization is difficult. Although earlier studies mostly depend on conventional machine learning and recurrent neural networks, we evaluate the efficiency of transformer-based models, which have not been extensively investigated for this job. In this work, for Arabic meter classification we investigate pretrained transformer models such as Arabic Bidirectional Encoder Representations from Transformers (Arabic-BERT), BERT base Arabic (AraBERT), Arabic Efficiently Learning an Encoder that Classifies Token Replacements Accurately (AraELECTRA), Computational Approaches to Modeling Arabic BERT (CAMeLBERT), Multi-dialect Arabic BERT (MARBERT), and modern Arabic BERT (ARBERT), and deep learning models like Bidirectional Long Short-Term Memory (BiLSTM), and Bidirectional Gated Recurrent Units (BiGRU). The half-verse data with 14 meters were employed in this study. The CAMeLBERT and BiLSTM model shows 91% accuracy compared to other models. We investigate feature significance and model behavior using a public dataset utilizing the Local Interpretable Model-agnostic Explanations (LIME) interpretability approach. These results show the benefits and constraints of every method, therefore opening the path for further developments in Arabic poetry analysis with deep learning.

Abstract: Open-domain event extraction, which aims to identify and structure event information from text without predefined schemas, remains a challenging task. Traditional methods often struggle with the diversity of real-world events, while recent efforts leveraging large language models (LLMs) show promise but still face challenges in effectively extracting structured information and inducing event patterns. In this paper, we propose a novel two-stage generative approach built entirely on LLMs. Our method first employs instruction tuning to train an LLM to generate natural language descriptions of events, including triggers and argument roles, from input text. Subsequently, we introduce a meta-learning inspired few-shot learning strategy that enables the LLM to implicitly learn event patterns and identify common argument roles based on the generated descriptions. We evaluate our approach on the ACE 2005 and ERE benchmark datasets, demonstrating significant improvements in F1 score compared to strong baseline methods, including traditional supervised models and other LLM-based approaches. Furthermore, ablation studies validate the contribution of each stage of our method, and human evaluations confirm the superior quality of the extracted event descriptions. Our work highlights the potential of a purely LLM-centric approach for flexible and effective open-domain event extraction and pattern induction.

Abstract: Objective: To present a framework for responsible implementation and evaluation of AI Agents in clinical service delivery, focusing on their potential to enhance healthcare efficiency, improve diagnostic accuracy, and personalize patient care.Materials and Methods: We outline a six-part framework for developing AI agents, including foundation model selection, adaptation for a healthcare domain, integration with third-party tools, hosting and infrastructure details, software stack design, data security and privacy considerations, and performance and evaluation.Results: We demonstrate our framework through an application on the example of a triage and scheduling AI agent developed for a hypothetical specialist medical clinic, illustrating key trade-offs and decisions throughout the system development and including illustrative code listings demonstrating how various system components come together in practice.Discussion: We highlight the transformative potential of AI agents in healthcare while addressing critical ethical considerations, including bias mitigation, transparency requirements, and patient privacy protection. Implementation challenges encompass technical barriers, organizational resistance, and regulatory compliance needs.Conclusion: The framework provides a comprehensive approach for healthcare institutions to implement AI agents effectively, demonstrating their potential to enhance clinical service delivery through improved efficiency, better decision support, and personalized patient care. We emphasize the need for continued research, collaborative data sharing, and supportive regulatory frameworks to advance AI integration in healthcare settings.

Abstract: E-commerce platforms use personalized recommendation systems to improve user experience and increase sales. Traditional models have difficulty capturing the complex relationships between user behavior and product features. This paper introduces the Enhanced Attention and Interaction Network (EAIN), a new approach that combines higher-order feature interactions and attention mechanisms. EAIN includes modules like the Dynamic Interest Network (DIN), Selective Feature Interaction (MaskBlock), and Position-Aware Interaction (PAIM), using data preprocessing and feature engineering to improve user-product relationships. Experimental results show that EAIN performs better than traditional models, especially due to the attention mechanism and higher-order feature interactions. This work contributes to personalized recommendation systems in e-commerce.